CSINT Conversations: Stopping online abuse of children - Could Apple have the answer?
Thank you all for being here for today's event. Where we're going to be discussing Apple's proposed CSAM policy and a bunch of other fun questions that are related to that new tech. So I’m going to go ahead and introduce our wonderful panel today starting with Laura Draper. She is a Senior Product Director at the Washington College of Law's Tech Law and Security Program where she manages research focused on identifying the tools, best practices, and legal and policy options available to tech companies and law enforcement as they combat online child sexual exploitation. Prior to joining the College of Law she served as an Assistant General Counsel with the Federal Bureau of Investigation and as a judicial clerk in the southern district of New York.
She also previously worked at the Council of State Governments Justice Center where she focused on law enforcement matters. She earned her BA from Case Western Reserve University, an MS from the University of Pennsylvania and an MPhil from Cambridge University, and a JD from NYU School of Law. We also today have John Verdi who is the Senior Vice President of Policy at the Future of Privacy, where he supervises the organization’s extensive policy portfolio on a broad range of issues, including: Artificial Intelligence & Machine Learning; Algorithmic Decision-Making; Ethics; the Internet of Things; De-Identification; and Drones and so many more other policies.
John previously served as Director of Privacy Initiatives at the National Telecommunications and Information Administration, where he crafted policy recommendations for the US Department of Commerce and President Obama regarding technology, trust, and innovation. Prior to this work, he was also General Counsel for the Electronic Privacy Information Center also known as EPIC, overseeing their litigation program. He earned a B.A. in Philosophy, Politics, and Law from SUNY-Binghamton and a J.D. from Harvard Law School. And last but certainly not least we have Dr. Hany Farid who is a Professor at the University of California, Berkeley
with a joint appointment in Electrical Engineering & Computer Sciences and the School of Information. He previously served as the Dean and Head of School for the UC Berkeley School of Information. Dr. Farid’s research focuses on digital forensics, forensic science, misinformation, image analysis, and human perception, and he has consulted for various intelligence agencies, courts, news organizations, and scientific journals seeking to authenticate the validity of images. He’s also the recipient the prestigious Sloan Fellowship and Guggenheim Fellowship for his work in the field, as well as a lifetime fellow of the National Academy of Inventors.
He earned an undergraduate degree in Computer Science and Applied Mathematics from the University of Rochester and a Ph.D. in Computer Science from the University of Pennsylvania. So thank you to all of our wonderful panelists who are here today. As a kind of intro to what we're going to talk about today - in August Apple announced a planned suite of new child safety features, including scanning users' iCloud Photos libraries for Child Sexual Abuse Material also known as CSAM, their Communication Safety to warn children and their parents when receiving or sending sexually explicit photos, and expanded CSAM guidance in Siri and Search. Following their announcement, the features were criticized by a wide range of individuals and organizations, including security and privacy researchers, the Electronic Frontier Foundation also known as EFF, politicians, various policy groups, and even some Apple employees themselves.
The majority of this criticism was leveled at Apple's planned on-device CSAM detection, claiming that it was dangerous technology that bordered on surveillance and in itself ineffective at identifying images of child sexual abuse. So, Apple actually went ahead with the Communication Safety features rollout in their Messages, which went live in December of last year, but Apple actually decided to delay the rollout of the other detection tech. But they have stated that they still intend to move forward with their detection technology, they're just taking some time to gather information and sort of think about that feedback from everybody who had comments and concerns. So to start us off, Laura could you give us a little bit of background information on what it is that we call CSAM? Sure. So, I think of CSAM as actually a subpart of a larger conversation around online child sexual
abuse and exploitation. So to the extent that that is our umbrella term, there is within that three primary patterns which are neither mutually exclusive nor exhaustive. The first of those is what we traditionally think of as when we think about this issue which is a child is abused in real life, images are taken of that abuse and uploaded to the internet and then shared among like-minded individuals. Those images are referred to as CSAM - child sexual abuse material. In the U.S.. code, criminal code, they are also referred to as child pornography. Most people who work in this space prefer this term CSAM because children cannot consent to their own abuse and so CSAM better captures the gravity of the situation. So that's our first kind of pattern of behavior. The second I would refer to as perceived self-generated content. So that's
children, that is people under the age of 18, who take photos of themselves in sexually compromised or explicit ways - those images are then shared knowingly or unknowingly by the child and in some instances they can then be the basis of sextortion which is where the recipient extorts the person who created the images in the first place to produce increasingly explicit images under threat of public release of those images. The third category of online child sexual abuse and exploitation that I think of is internet facilitated prostitution of children, which follows what we would typically think of as the pattern of prostitution of children, but that is aided by the internet in some way or fashion through either communication technology or through online advertisements and other things of that sort. So those are really the three primary patterns. Of course, there's overlap. They're not exhaustive. There are things I’ve not mentioned, but so we can think about both the umbrella and then more narrowly into the CSAM. Great. Thank you so much for that overview. I think obviously we have Hany on the panel today,
who perhaps is our technical expert here with all those computer science degrees and so maybe you could start us off and kind of get us going on thinking how effective would Apple's proposed photo checking tech strategy really be given the current use of this technology to distribute the content as Laura had mentioned? Good, thank you Divya. So there's… if we acknowledge that we would like to develop and deploy technology to limit the spread of this - so let's put aside the privacy argument and all the arguments that I’m sure we will be having, just put that aside for a minute - there's sort of two problems if you will that you could try to address. One is - for every image, every video that is uploaded or sent, try to determine if there's a child in the image or video, if it's sexually explicit, and if the child is underage - that's a really, really hard problem. Even with all of the advances of today of machine learning, of AI, the technology is not there to work with the accuracy and speed to analyze literally billions of images and video a day. So we don't have that artificial intelligence ability to reason about the underlying content the way our human visual system is capable of doing. Now another way to think about this problem - and for this I have to explain one thing that you should understand about CSAM - which Laura did a very nice job of giving us some scaffolding to think about - is that the same content is distributed year after year and decade after decade. So once the content is created, the original exploitation of a child
and that exploitation is recorded - that content lives on - for in many cases decades, and in many cases that image is at some point identified by law enforcement, the National Center for Missing and Exploited Children, the Canadian Center for Child Protection, Interpol, whoever it is. And so a different way to think about this problem of limiting the spread of CSAM is to say we have previously identified these pieces of material. Today at the National Center for Missing and Exploited Children that number is in the tens of million and - we are going to limit the spread, the redistribution, so for content that is newly created that we have not previously seen, we will be blind to it, but the reason we think about this problem is that technically that problem is a little bit easier. Because I’m handing you an image or a video and I’m saying
this has been identified as a child underage and we know it is sexually explicit, we know it it's categorized as CSAM - we would like to stop the redistribution. And that's the category that we are talking with the Apple technology and photo DNA, which is the technology I worked on with Microsoft many years ago. A couple things I want to say about this. So one is, it is technically feasible. We can scan images with very very high accuracy and very very high
efficiency to determine if it is previously identified CSAM and I think the most important thing to understand here is that when we develop photo DNA, which is this type of technology for stopping redistribution - it was never meant to be a law enforcement effort. It was meant - it was a victim-centric technology - because if you talk to the children who are abused what they will tell you is that was probably the worst day of their life, but what they'll also tell you is day in and day out knowing that the images in the video of their abuse being shared online is horrific for them. So if we can disrupt that global distribution of past abuses it is very much in the interest of the children and in their interest - and in their privacy as well. And so, very briefly the way the technology works is it extracts from a piece of content an image or a video, a distinct digital signature, and it tries to match that signature on upload. That's how photo DNA work. That's the core of the way Apple did it. Apple
had some nice features that were more privacy preserving - which ironically got them in trouble with the privacy groups even though it was the more privacy preserving technology than photo DNA. I’m sure we'll talk about that in more detail. And the last thing I’ll say about this is that this technology is very similar to other types of cyber security technology.
Every time you send an email every attachment is compared against the signature of malware, viruses, spam filtering…every time you upload a file those are scanned against known harmful content. So that mechanism, that scaffolding that Apple is using, that we have used in the past - sort of, they already exist in cyber security. We're not really getting very far out of accepted best practices. Now how effective is it? It's a great question. Is it going to stop the abuse of children? No, of course not. Does it protect children? Absolutely. Just to give you an idea. Last year alone - in one year the National Center for Missing and Exploited Children - their cyber tip line received 20 million reports of CSAM - the vast majority of which were done with photo DNA, this type of technology that we're talking about. That's 20 million images of kids being
abused that did not get distributed. Does that solve the problem in its entirety? Of course not, but nothing is going to solve the problem its entirety. The way you solve this problem is you chip away at it - and I consider this type of technology the absolute bare minim that we should be doing to protect children online and to protect their privacy. Thank you so much for that. I will say before I move on, that I had meant I meant to mention
this in the beginning, but if you do have any questions for our panelists, please be sure to write them in the Q and A section and we'll definitely have time at the end of the panel to have our panelists answer your questions. So, building off of what you've said, obviously you've mentioned that privacy advocates have had issues with what Apple’s put forth. There was this idea that Apple would integrate privacy features but of course perhaps not to the liking - or to the level - that privacy advocates would like and so I’m going to ask now, John, you know - was that - is there really valid concern about the privacy of people's personal sensitive photos with the tech and that Apple has proposed? Sure, so let me first say that my views on this matter are probably more aligned with my fellow panelists than some others in the privacy community. And I very much appreciate the history
on photo DNA and I appreciate the context here. I’m actually sort of glancing down at my notes and my colleagues here have made some of the points that I was actually going to tee up to sort of frame this technology. So thank you so much for doing that. That gives me the fun part, right? Where I can delve into the controversy. So let me say this - the concerns that have been expressed by cyber security experts and privacy experts about Apple's proposed technology are no doubt real. The question in my view and I hope we can drill down on this, is whether or
not those concerns are greater, equal to, or less than alternative models. Models that are used by industry every day today, not by Apple, but by others in industry - other big players, right? Communications companies, cloud storage companies, etc. in order to detect and mitigate this terrible material. And I think that's a really robust conversation that I would love to have. So in order to perhaps tee up some of that conversation, let me just flag a couple of things. I appreciate the kind introductions - the one thing that I have in my notes that we have not covered so far is the legal status of this material - and I think it probably goes without saying for my colleagues up here - but I just want to flag it for attendees and for anybody who's watching us on recording - CSAM material is unique when it comes to moderation and law enforcement and the desire of platforms to have this material off of their servers. Unlike other material - even other unlawful material - CSAM material is in many jurisdictions a strict liability offense.
That is no intent is required in order to prosecute. There are regional differences, but anyone who would say that the legal status of CSAM is equivalent to the legal status of say - material that infringes copyright - does not understand this issue. Right? Intent matters when it comes to copyright infringement. Fair use analysis matters when it comes to copyright infringement. Commercial versus non-commercial use matters when it comes to copyright infringement.
There's lots of different things that matter there. CSAM material is illegal to possess. It's illegal to create. It's illegal to distribute. It's illegal full stop. And there are lots of good reasons beyond the legal reasons for why platforms want this stuff off their systems. So if you're a platform you have a few choices. You can do nothing - which virtually assures that your platform will become a haven for this unlawful and terrible material. You can scan content on your platform - which is
effective to some degree, as my colleagues have described, but necessarily requires that you as the platform developer and owner have access to all that material on your platform - in the clear. Or - and this the novelty of Apple's approach you can do pre-upload comparison. Much in the way that services do pre-upload comparison for malware, for hacked data, right? Data that has been released, that has been made publicly available in a data breach and others try to upload and distribute on cloud platforms. There are pre-upload checks for that. Some do pre-upload checks for copyrighted material. Some do pre-upload checks for material that is not technically speaking malware, but can
be problematic, such as bitcoin mining software or crypto mining software that can take up a lot of computer cycles and a lot of energy and it's against terms of service, right? So this kind of pre-upload matching that Apple has developed poses different questions from the privacy community than on-platform scanning does. And in order to tee up what I think is the really interesting discussion here, of how does this compare to other approaches sort of doing nothing - I think it's important to appreciate that the critiques from the cyber security community and the privacy community have in my view largely fallen into two buckets. One bucket is Apple's approach to pre-upload matching of CSAM vulnerable to abuse by governments or other bad actors who would like to turn this tool into something for which it was never intended, and instead use it to try to identify other sorts of images, non-CSAM images. That's one bucket of concerns. Is the actual - to Apple approach and technical implementation vulnerable to attack if you're a cyber security professional. The second bucket of critiques has been more of a legal and policy and global political critique. And that has been
– this is the essentially the first instance of pre-upload matching on Apple devices for photo storage and will this encourage governments around the world, law enforcement, agencies, corporations, individuals, whomever to press for increased surveillance and pre-upload scanning and matching regarding other material. Not a technical attack on the Apple implementation, but a legal either requirement or request or incentives to expand this beyond CSAM. Those are the - in my view - the two buckets of critiques. And those are the concerns that were expressed by cyber security professionals, by privacy professionals, and I will say that the manner in which Apple rolled out this feature - regardless of the substance of the feature, right, perhaps could have benefited from pre-roll-out vetting by independent security professionals. There was, I don't - I want to be clear - there was some pre-rollout vetting for… by select experts, but I think a broader vetting might have made a difference in the reception - or perhaps not. You bring up some great, great questions. One of which actually was one that I had
myself for all of you to answer - and you know, I’m happy to hear also what our other panels have to say to John's questions. that specifically - I had the question you know, how - you know, what the valid risk is of abuse by other countries, where personal freedoms and civil liberties are really at high risk of persecution by the government. And so I open this up you know, to Laura and Hany to provide some insight as well. 20:31 Sure, so I would say, as to John's two questions - the first question being is Apple's system as set up, subject to abuse by other governments - and the second question is does Apple's system open up to broader abuse, mandating other systems to have that kind of on-device scanning. So,
with respect to the first one, I actually think Apple has been quite clever and deliberate in how they've structured this. They require the scanning, actually is, gets matched against databases of these CSAM fingerprints, so to speak, on these hash values and it has to have, in order to be a positive match, it has to hit against databases owned by and maintained by two sovereign jurisdictions. So that is to say that the United States could not make the decision to say that some terrorist video should be identified and rooted out and so we hash it, put it in our database, and then put it in our - another one of the US’s databases and then we get two matches and it gets uploaded. It's two sovereign jurisdictions. So that itself actually is quite protective I think in a lot of ways. In terms of in the ways that governments could or could not manipulate the existing Apple system. Whether or not it opens the door as to the second question is a little bit of a stickier issue, I think in some ways, but the first one I think, like I said, I think Apple has actually been quite deliberate in ensuring that the way they structure it would be pretty protective in that sense and pretty resistant to abuse.
I’ll say a few things. I think John did a great job of laying the landscape of the concerns that people have and I think John is absolutely right that there are legitimate concerns here. We should - and we should talk about them. So is it vulnerable to the so-called slippery slope. Well if you do this what are you going to do next? And you know that argument I always find a little intellectually lazy because you could say that about every technology. So for example there is a microphone on my phone, there's a GPS tracking could - that be abused? Of course. Do
I know what's inside of the IOS every line of code - how that information is being used? No. Could malware detection, spam detection, virus detection - any of number of cyber security technology be abused by companies, by hackers, or by repressive regimes? Of course. So if we're going to use the argument “could this technology be abused” - we all need to throw away our phones and turn off our computers and unplug them, because all of the technologies can be abused. And so the question is not can it be abused - and Laura did a nice job and did a good service for us is explaining that Apple actually was very thoughtful in, and I would say they actually went too far in some cases I think they actually made it too hard, but I appreciate why they did it. It's extremely privacy preserving. They put the safeguards in place. Is it perfect? Is it open to
abuse? Of course, but then as John was saying we have to balance that. What are we talking about on the other side of that and what we're talking about and what we haven't mentioned by the way - and I think it's worth mentioning without being too graphic - is that the average age of a child involved in CSAM is eight - down to a few months old. We are actually not talking about 15, 16, 17 year olds who are playing with their sexuality. I’m not excusing that, don't get me wrong, we are talking about pre-verbal infants who are being sexually abused and their contents being distributed online and I say that because you have to put the risks and the concerns on balance with those issues of protecting the most vulnerable among us and I would say this as I said earlier - the bare minimum technology. And I want to also emphasize that what I found really interesting about the pushback, was Apple did, is doing, proposed to do the scanning on the device for images that were about to be uploaded to the cloud - and that's what made people crazy. If
they had said “hey, we are going to scan images as they enter the cloud” people would say, “well that sounds about right”. Everything gets scanned when it enters into the cloud for all forms of abuse, as John enumerated. But doing that means you can't store things in an encrypted fashion, so to preserve user privacy we are going to incorporate this technology on your device - which is the right thing to do for a privacy preserving technology, but it is exactly what made the privacy groups extremely unhappy. And I think primarily, as John enumerated, because - well what's next? And the thing that I have not found compelling is - when you when you make a slippery slope argument you have to ask at least two questions. One - is it possible? And the answer
is yes. And the other question is what's the incentive? Like what's the incentive for Apple to actually do something that is nefarious - and that is missing here. So I don't find those threats on balance outweighs the benefits of the technology to protect children online.
So can I add on to that very very briefly? Definitely go ahead. So, I don't want to - I don't want to step on our kind moderator and friend, but you can just because you referenced your development of photo DNA and you've been at this for a while I’ve worked in this space for a while too and one thing that I think is helpful is to have an appreciation for the history of how these technologies have been implemented over time. How NCMEC has operated over time as a trusted repository of these hashes, right? And it's not just the technical controls that Apple has put in place. It's not just the process and policy controls in terms of human review that Apple has put in place but when one talks about the risk in my second bucket - which I call the political and policy risk. And to be clear I am very very strongly supportive of strong encryption and encryption technology. Right? Go to my twitter, check out the resources The Future of Privacy Forum produces. I am probably further over
on being supportive of end to end encryption than a lot of people in the field for a lot of different reasons, right? But one thing that I think is helpful here to understand and appreciate is that we have a decade plus experience of how NCMEC operates, how it partners with companies, how it interacts with law enforcement, right? How other similarly situated entities operate and we have a decade plus of experience in terms of at least having a track record for whether or not steps to identify CSAM and mitigate and eliminate CSAM lead to government overreach in other content areas. There are a lot of negative stories to tell about privacy out there in the world in the commercial space, in the law enforcement space, in the government space, generally in the U.S. and around the world, but I don't consider overreach regarding CSAM mitigation to be one of those stories. We simply don't have a political track record for a decade plus of lawmakers and regulators and law enforcement agencies around the world successfully compelling global companies to expand their photo DNA efforts to include copyrighted material, politically objectional material, other sorts of material that governments and law enforcement agencies and others around the world would love to curtail, right? We simply don't have a track record of successful attempts there. So when we talk about the slippery slope, I think it's important to ask based on our knowledge about the world and the political process, how slippery is this particular slope? I am not here to tell you there is a 100% guarantee that implementation of this feature would not lead to renewed calls in legislatures and in regulatory bodies in the U.S. and around the world for expanded scope. I am here to say, that we have a
track record on these issues and it's not a track record of a year or two years it's track record of a significant period of time in which the - in my view - common sense mitigation of CSAM, using photo DNA and other similar technologies, has not led to that sort of overreach. I don't know Hany if you have anything else to follow up with that. I really appreciate that John. I think you're absolutely right and I’ll just say… So photo DNA was developed by myself and Microsoft back in 2008, so John is right, it's over a decade now. And I will tell you, when we started deploying it in 2009 I heard almost verbatim the exact same objections that I’m hearing today with Apple's technology - this going to lead to xyz and that, as John said, has simply not been true. And I think we have to acknowledge that John is absolutely right. Is it
100 guaranteed, no. But again we have to balance what are we trying to do on the other side of that and I think there have been - we put reasonable safeguards in when we developed and deployed photo DNA. I think Apple has done the same thing. I don't think anybody wants to overreach here, but I think that you know, we are in crisis mode about protecting children online and I will say things have gotten much much worse over the last two years through the pandemic. The number of child sexual abuse imagery that is being reported every year keeps going up. We now have live streaming of child abuse on platforms like Zoom and other types of video conferencing and the problem is getting worse and we have to acknowledge that some interventions are necessary. And the thing
I’ll say too - is I’m always frustrated that you know, I’ve heard people say - well it's fine to put technologies on my device to protect me the user – malware, virus, spam, etc. etc. but don't you dare put technology to protect other people. And I think that is awfully selfish and we should acknowledge that yes, we are giving up something - I think it's very very small, but it's for a very very good reason that we're doing that too. Laura, I don't know if you have anything else to add here. Obviously they're a very complicated debate. It is very complicated. Just to Hany’s last point - I think you know, the idea of these trade-offs - you know don't infringe my privacy even for the protection of children. I think one of the things that gets a little
bit lost sometimes when we talk about privacy as this kind of monolith term, we're talking about the privacy of us - you know frankly, us on this panel – right? The law abiding citizen right, but the privacy interest that often gets overlooked and even lost in these conversations is the privacy interest of the abuse victim of the child - who has been abused and so they also have a privacy interest in not having their image of, as Hany said, the worst day of their life most likely, continued to be circulated and distributed and shared. And so I think that also just needs to be incorporated into when we think about balancing these issues. No - it's a it's a valid point, obviously. I think as everyone knows once you're on the internet you're on the internet. There is no removing that content once it's out there in the world and so I, you know we have been talking about this sort of photo detection technology and I want to pivot a little bit and also talk about the communication safety feature too that Apple has rolled out. And Laura I know you do some research as well on end-to-end encryption and how that plays a role in this and so I pose this question to you about you know, we have this constant debate about access to and encrypted messages - is Apple's message check tech and strategy really a step in the right direction or the wrong direction? So, I think it's a step in the right direction and I think that for a handful of reasons.
First of all, I think end-to-end encryption is here to stay. I think we can have as many debates as we want about it and law enforcement can scream from every top mountain and hill and it just doesn't matter. End to end encryption is going to be a thing and it's already here and it's only going to become more prevalent I think across all of our platforms. So with that in mind we
have to come up with strategies and solutions that combat this issue at every way we can. We have to take every bite at the problem we can. As we talked about earlier, there's a whole range of harms that occur against children online in these contexts and so we have to think about multiple strategies. There is no one fix to this problem and so in thinking about this as well,
just to give some numbers - you know Hany pointed out the cyber tip line that gets the reports received by NCMEC, the National Center for Missing and Exploited Children, in 2020 there were over 21 million reports made to NCMEC. Facebook made over 20 million reports. Apple made 265. Not thousand. 265. And so Apple needs to be - and is taking this issue seriously. They are stepping up their work and every bite that they can take at this apple is a step in the right direction.
When we talk specifically about the iMessage issue again, as Hany pointed out, with respect to the tech when we think about the algorithms that have to identify this, all they're really doing right now is identifying nudity - and that's for technical reasons I suspect as much as anything else, in that it's easy to identify nudity. It's much harder for a machine to identify CSAM and so I want to be sure that those two things don't get conflated, but within the realm specifically of, in a space like iMessage where it's end-to-end encrypted, you've got this machine learning to identify sensitive or explicit content. I think one of the things that's quite useful in this that it doesn't actually prevent the sending or receipt of the material all it does is prompt the user to say “are you sure you want to look at this?” “are you sure you want to send this” “are you really sure that you want to send this” - and that kind of like road bump so to speak is actually quite useful and provides the user with a great deal of agency about what they do and do not want to view. It is not going to stop the adversarial distribution of CSAM. It's not going to stop bad guy to bad guy distribution or sharing of content. It's not, but as we've talked about - this perceived self-generated material
is very much on the rise and having that kind of prompting to remind users that this pic - this image is going to be shared with somebody - I think can be a useful pause in the way young people and children think. The Internet Watch Foundation which is a non-profit based out of the UK reported that, I believe, in 2020 or 2021, 44% of the reported images they received involved self-generated content so this a very much a big and growing problem and the pandemic has of course certainly made it worse in many respects. So it is something. And again, so you're tackling a problem that really is a kind of the core and is growing and I think anything we can do to chip away and minimize the harms that arise from this a good thing. I would also just say that I would want to
be sure that it is coupled with a couple things. I think first of all it has to be coupled with preventative education to children, so that they understand what it means to be sending this content these images of themselves - what the risks are you know? Is this really a person you trust? When we think about sextortion Thorn, which is a nonprofit based here in the United States around online child safety issues, reports that about 60% of sextortion victims knew their the offender in person - in real life. And about 40% of them met and only and engaged on the internet. And so there really has to be a combination to say you might know this person in real life that doesn't mean that they're not at risk of doing something untoward with your image.
And so that - and I also think it needs to be coupled with a safe and secure user reporting mechanism within the program, so that if somebody is receiving or being prompted to send images that make them uncomfortable that they feel safe and secure reporting that to the platform or law enforcement or whomever - and that they understand that they have options. Hany and John, if you have anything else to add? And I’ll also say, you know we'll be waiting at the end for questions, but someone did ask the 265 reports that were made from Apple - was that a result of the complaints made to Apple or was it Apple doing anything proactive themselves? So I would say - I’m not totally sure I’m going to bet that it's based on user reporting as much as anything else, but the numbers are not broken down that way on the NCMEC site. Oh go ahead. Go ahead John. No please John. I’m not going to speak for Apple either, but I
would say just looking at the number my analysis is the same as Laura's - it appears to be as a result of reporting and not as a result of scanning. I will say, and I’m not going to attribute to a particular person, but folks who are in a position to know in industry or folks who previously were in industry in very senior roles and are now in other roles outside - have essentially characterized the volume question on this in differences between platforms - as largely based on the platform's willingness and volume of scanning, rather than a difference in volume on particular platforms. Which was just - which was surprising to me frankly. Other material - there are certain platforms that are particularly popular for posting music and particularly popular for posting movies and particularly possibly popular for posting documents and they don't always translate across platforms, right? One does not go to Band Camp to find a word document right? You go to find audio. But
what has been shared with me from people, again who are in a position to know, is essentially that many many platforms suffer from this problem and are actively combating it and that when you see differences in reporting numbers it typically has to do with the scope at which they are scanning, rather than the actual prevalence on their platform and that means that companies like Facebook who scan a lot end up in some circumstances getting criticized for reporting a lot even though it's simply a result of them doing more scanning. Hany, I don't know if you have a view. I think John you're absolutely right. And also we have to remember Facebook is Instagram, Facebook is WhatsApp, and so that's a very very big entity worldwide and I think John is absolutely right. The more you look the more you find and that's what we've learned over the last decade. I’ll add one thing I want to come back to what you said, just so that John and I don't agree on everything. I want to come to the question about end to end.
I’m you know, I’m not a firm believer in end to end and here's why. First of all, in the offline world there is nothing that is immune to a lawful warrant, right? My mail - my physical mail communication - my phone calls, my physical body, my office, my car, my home, my office - with a lawful warrant can all be searched and there's a reason for that, right? There's a reason why we do that and so I’m not I’m not entirely clear why we should have bulletproof end-to-end communication online that is immune to any type of security or lawful warrant access. I understand, don't get me wrong, that there are some bad people out there. There are bad governments. There's a
need for privacy and there's a need for secure communication. I absolutely understand that, but in the same breath we have to acknowledge there are really bad people out there doing bad things within this encrypted service - whether that is the sale of illegal drugs, which in the United States opioid deaths last year topped a hundred thousand - vast majority of the dealing of those drugs is happening online -child sexual abuse, terrorism and extremism, the on…the sex trade and so there are really bad things happening. And to create this bulletproof tunnel that allows us to do that. I’m not entirely sure that the benefits of that outweigh the drawbacks and again, it's all - I see the benefits, I really do, but we have to acknowledge - and even Zuckerberg acknowledged there is a threat if we deploy end-to-end encryption and I’m not there yet. I’m not sure that I see the benefits outweighing the drawbacks and I see what Apple did is a very thoughtful entryway into trying to respect and preserve the end to end encrypted messaging service while adding some protection. And maybe Laura is right,
is that it is here to stay and we just have to figure out how to manage it as opposed to trying to eliminate it, but I think I’m not - again I just want to say - I think there's still a debate to be had as to whether the benefits are going to outweigh the drawbacks. I’m on hold. John do you care to respond? Good. Now we disagree about that. Now we're getting into it. Now we're getting into it. So on end and generally my - I don't disagree with the global analysis that end to end has benefits and drawbacks. No question. What I do disagree with is whether or not this a fait accompli and the game is already over. And I would say my view is that it is,
end-to-end encrypted messaging storage etc is technically trivial to develop. Criminals of all stripes will have it whether governments around the world want them to or not. The only question is whether the rest of us, the non-criminals - and I’ll assume for the sake of argument today that the panel and our guests are non-criminals, at least in a meaningful way - the question is whether or not we derive and society overall derives the benefits from end-to-end encryption because criminals will, whether we want them to or not. Okay, so let me pose this. Here in the U.S., we have I think the world's largest number of handgun
and gun deaths by a spectacular margin. And can I make the same argument here? Bad guys have guns, so let's give guns to everybody, what could possibly go wrong? So is that the same argument John? By the way I appreciate the argument - all right - that's one of the most thoughtful arguments. But is that the same argument? Well bad guys have guns, well then I want guns too? If guns could be independently manufactured and developed by a smart undergraduate student in a weekend in their dorm room, maybe. I see - you're saying – yeah. Although 3D printing of guns is sort of getting us there, but I don't think we're quite there yet.
But genuinely… I see what you're saying is. I think this is a principled position that I believe. That the facts about the world are that the underlying math and the underlying software implementations are out there and exist and if you asked me the question “hey, we have a lot of bow and arrow deaths here in the United States” - which I’m not comparing to guns in a meaningful way, but a bow and arrow is very simple to create oneself, right? You know what we regulate bow and arrows - when someone can simply go into their backyard and take down a sapling and create a bow and create an arrow – yeah, I look - I’m happy to have the overall conversation about costs and benefits and everything else. My only question is, given the open source libraries that are out there, given the underlying math, given the other underlying technical implementations - my sense is, I know you said you're not there yet, my sense is that I am there yet and these are the same technologies, just to be clear when you talk about end-to-end messaging, that the Department of Defense recommends that U.S. soldiers abroad use to communicate with their families back at home.
Yeah I get that. I get that. Let me add one more thing and I don't want to hijack the conversation - let me add one more thing - is that- there is still, you're absolutely right that you know you can go to GitHub and download the technology for creating end to end, but there are still gatekeepers here right? Because if you want that app to be used there is still the Apple store, there is still the Google store, where these apps have to be deployed and though and those stores routinely ban apps that you know for various types of reasons. So there's still a little bit of a gatekeeper there. It's not just “hey I can deploy this and have a messaging with
my friend”, like there is still a barrier to entry here in terms of getting wide deployment. Yeah. Yes I agree there are there are gatekeepers and we can go back on this, but I do think we've identified the nub of the of the disagreement. Okay, good. I would also add you know to this conversation, we've - you mentioned you know, there are victims who are out there. And of course from a parent's perspective of child, whether they have been a victim or not, most likely would advocate for yes, please anything that will protect my child, please put it out there – regardless. So I’m curious to know - you know, do those people then outweigh whatever other kind of conversation to be had when it comes to this subject matter and saying that yes we want more and more protection? I don't know. I don't know how to balance these things. And I can tell you I’m not the right person to ask because I have spent the last more than decade you know talking to young victims of these crimes and it is gut-wrenching and it is it is heartbreaking and it happens on a daily basis and it absolutely - I will admit colors the way I think about these issues in a way that makes me I think not particularly objective.
And I think it's you know - I challenge anybody to go talk to some of these young victims and then take a strong position to counter what they say, but I don't think that's actually the right way to argue I think we have to think more on balance - because John is absolutely right that there are benefits to this technology and I don't know how to weigh those against some of the horrific things that are happening to kids around the world. I generally don't know how to weigh through them Laura do you have any comments on what we're talking about? Nothing additive. More than sufficiently covered by my fellow panelists. Well then I’ll go get to kind of the final question for today. You know we've talked about Apple has proposed so much detection tech and all these different strategies and we kind of come to this overarching question of really should we be leaving it to these tech companies to kind of determine how it is that our personal information and sensitive photos and things like that should be shared and scanned and you know detected and told to law enforcement or the government in some form - or does it require external regulation by some sort of a government entity to then tell tech companies how they should be leading forward? So I would say it should not be left strictly to tech companies. I think in an ideal world congress would be passing some sort of law, about not necessarily to say you have to use this technology or you have to scan in these ways not that level of minutia, but this the kind of personal data you should be collecting on people and should be able to provide to law enforcement upon legal request. One of the parts of my research that I’ve been doing over the past few months is that I’ve been talking to a lot of law enforcement officials around how they investigate and pursue these types of offenses and one of the primary complaints that I’ve really heard from them is an inconsistency across the way companies handle this. So what they have basically said is that every company - every
tech company - has a different policy for receiving legal process - which is to say subpoenas and search warrants. That they maintain different types of information about their users for different lengths of time. And when they make their initial tips to NCMEC that they include different categories of information in those initial reports. That kind of knowing what to ask for, how to ask for it, when to ask for it, and how to frame it to which company at which time, creates an enormous administrative burden for law enforcement officials investigating these sorts of offenses. And so when we talk about the fact that there are
you know over 20 million reports to NCMEC in a given year, that is itself a huge ask of law enforcement to investigate these offenses. And yes some of them is old repeated material and we know who the victims are and they're not being harmed physically harmed anymore, but that's not always the case and so anything we can do to alleviate some of the administrative pressure on law enforcement I think is to our societal benefit. And I think that having some sort of industry standard would be useful in terms of that types of - that type of thing, but like I said I would not suggest that congress get into the weeds about this technology in these ways and these circumstances. I think that's not going to be nimble enough frankly
or agile enough and would probably not keep up with technology in any meaningful sense. And I open this up also to John and Hany for your thoughts on it as well. Yeah I mean, I would simply say I appreciate the frustrations that the law enforcement officials Laura has been chatting with, you know have. I think what's interesting as we talk about iterating on the existing system and on the existing framework is that some of those frustrations about inconsistency are actually features, not bugs, in the original implementation of how folks got companies to sign on to do this reporting at all, right? By saying to folks “it's not going to be a prescriptive top-down reporting system”, “yes, you can still have your own retention periods, you can still have your own stuff and still participate and still make reports”, “yes, maybe you collect different information than your peer company, but as long as you provide account information to NCMEC then in the report that's okay”. The question is as that this framework has evolved, is there an
opportunity for standardization and normalization across those services now that we've onboarded folks over a decade plus? I think that's a really interesting question, but it's interesting to me. And I know you folks know the history of this in terms of getting folks on board at all to do this reporting there had to be a flexibility in order to do that and maybe now comes the harmonization. I’m not sure I like my choices - my choices between Facebook/Zuckerberg, or members of congress to make sure we are safe. This is not a great choice.
I think Laura got it right though and I think what we want is congress to put guardrails. I don't think we want a highly prescriptive you must do a b c d e f g down to z but I think some guard rails would be good. I think for the last 20 years there have been no guard rails or very very very few and I think we need to start putting in some guardrails around child safety, terrorism, disinformation, and all the various harm and privacy concerns that we have as well. I think we need to start putting in some guardrails. The thing that I worry about now is we already have virtual monopolies. We have a handful of companies out there that dominate. Trillion dollar companies that dominate - and one of the tricky things about regulating now is they are going to have a huge advantage over the new companies that are trying to come up with a regulatory landscape, that the current giants of the tech industry didn't have to deal with so we want to be very thoughtful that we don't start regulating in a way that stifles even more innovation than is already being stifled because of virtual monopolies, but I think if you look around the world here in Capitol Hill, the UK, Brussels, Australia, Canada everybody is trying to figure out how do we reign in the technology sector while - to mitigate the harms that we've been talking about - while allowing for an open free internet and for the things frankly that we were promised 20 years ago, but that I think have not really delivered. The internet today is
not the internet that I was promised 20 years ago. And I think I still, despite my frustration with the technology sector, I still think technology is a force for good, but I think we have to be more proactive in making sure that they're that the harms are now out - now are not outweighing the good parts right now, which I think is what's happening online right now. Wow, well thank you all for a really really interesting discussion. I can't believe it's already been an hour and there's clearly so much more we could talk about, but I do want to get to some of the questions that the audience has submitted. We already have a good number and so
one of our first questions is talking about the detection technology, which is called neural hash and they mentioned that it triggers a report once it's found 30 files in an event to attempt to reduce false positives, which is like saying you need to find 30 pieces of DNA at a real-world crime scene before an investigation starts. So could there be an argument that this overly cautious and neural hash should be more sensitive to detecting this type of content? Yes. I was not a fan of that. I think Apple probably went too far. There - here's my guess… I don't know why they did that, but here's my guess. If you assume that every image you scan is independent of the next image then the false alarm, let's say for a single image is one in a million, well then the false alarm of two images is one in a million times one in a million, and the false alarm for three images is etcetera etcetera - and so by the time you get to 30 it is an astronomically small number and I think probably an absurdly small number. Just to give you some context when we were developing photo DNA our false alarm rate was on the order of one in 50 billion - so that means you'd have to see 50 billion images before you had a false flag. And please understand when there's a false flag people don't go to jail, right? It's just a human moderator steps in looks at the image and makes a determination, so it's not a huge lift. I think that was probably overly cautious and I
would like to see that number come down to a much more modest number like five, let's say. But the question is correct. It’s like you're basically giving people a bye for the first 30 pieces of CSAM and that doesn't really sit very well with a lot of us I think. I’m guessing John and Laura you agree from the vigorous head shaking I saw? Yeah, I think also you know, I think - and I don't know that there are statistics on this one way or the other, but I think a lot of law enforcement officers who work on these sorts of investigations would tell you that it was just one piece of known imagery that resulted in them identifying an offender who was actively abusing children. And so you know, when you take that into account and we're talking about live victims who are currently being harmed, I think setting that kind of threshold as Hany said is perhaps too much. So I will jump in and agree to disagree on that this one. So I was nodding vigorously because I
agree with 99% of Hany’s framing and I agree with 99% of Laura's framing. Here's what I would flag. The idea that this gives a free pass to folks for their first 30 images. I get it. I completely get and understand the point and I take the point, right? Should the number be 30? Should it be 5? Should it be 8? Should it be 12, right? I think we should have a healthy conversation about that and I think that conversation should be informed by data, right? The data from Laura's work about law enforcement officers who use a single piece of known CSAM to identify an active abuser. That should be taken into consideration. We should also take into consideration though
what the spread of the data looks like when we have existing data on CSAM reports at companies who scan in the cloud today, right? What is the average number of offending CSAM photos or videos when a user is reported is it one? Is it 100? Is it 1,000? Based on some conversations I’ve had with folks, those numbers tend to be really high. Folks tend to hoard this material, right? Now I’m not - I don't have a clear view into this, so I take Laura's research and her work with officers very seriously and Laura says “well one is often a key to cracking these cases” - fair enough, right? I would share one case that was out of Texas in which a large cloud company which scans its servers all the time using photo DNA to try to identify known CSAM - they are you know a household name, you know, for folks in the tech industry. They are a good actor in this world and they make tens of thousands hundreds of thousands of reports every year - they made a report to a police department in Texas regarding a single piece of CSAM that was detected on an upload to their servers. The defendant in that case was a young man who stated he was in a video game chat on a well-known video game platform and someone pasted a url into the chat and said go here for more texture packs and plugins for this game. He clicked on it got an error page. Oh interesting. Okay. Turns out that that error page was masking a surreptitious
download of known CSAM and that surreptitious download of known CSAM ended up in his downloads folder on his computer and that downloads folder was set to automatically back up to the cloud. Everyone involved in this particular case agrees that it is a single piece of CSAM, that as I said before on a strict liability basis, this individual was no doubt in possession of unlawful material, but I think what Apple is trying to do is turn the dial a little bit to avoid situations like that. Not to give this individual a pass, but to separate that situation from an individual who actively seeks out this material and downloads five or 15 or 50 or 500 pieces of it - so that's just to articulate a little bit of the balance where I think they're trying to get at here. If we had a law about CSAM and I’m not recommending this at all that had an intent requirement, I think Apple would be far less worried about the scenario I described. But since it's rightfully a strict liability regime