Category Archives: Security engineering

Bad security, good security, case studies, lessons learned

Is Apple’s NeuralMatch searching for abuse, or for people?

Apple stunned the tech industry on Thursday by announcing that the next version of iOS and macOS will contain a neural network to scan photos for sex abuse. Each photo will get an encrypted ‘safety voucher’ saying whether or not it’s suspect, and if more than about ten suspect photos are backed up to iCloud, then a clever cryptographic scheme will unlock the keys used to encrypt them. Apple staff or contractors can then look at the suspect photos and report them.

We’re told that the neural network was trained on 200,000 images of child sex abuse provided by the US National Center for Missing and Exploited Children. Neural networks are good at spotting images “similar” to those in their training set, and people unfamiliar with machine learning may assume that Apple’s network will recognise criminal acts. The police might even be happy if it recognises a sofa on which a number of acts took place. (You might be less happy, if you own a similar sofa.) Then again, it might learn to recognise naked children, and flag up a snap of your three-year-old child on the beach. So what the new software in your iPhone actually recognises is really important.

Now the neural network described in Apple’s documentation appears very similar to the networks used in face recognition (hat tip to Nicko van Someren for spotting this). So it seems a fair bet that the new software will recognise people whose faces appear in the abuse dataset on which it was trained.

So what will happen when someone’s iPhone flags ten pictures as suspect, and the Apple contractor who looks at them sees an adult with their clothes on? There’s a real chance that they’re either a criminal or a witness, so they’ll have to be reported to the police. In the case of a survivor who was victimised ten or twenty years ago, and whose pictures still circulate in the underground, this could mean traumatic secondary victimisation. It might even be their twin sibling, or a genuine false positive in the form of someone who just looks very much like them. What processes will Apple use to manage this? Not all US police forces are known for their sensitivity, particularly towards minority suspects.

But that’s just the beginning. Apple’s algorithm, NeuralMatch, stores a fingerprint of each image in its training set as a short string called a NeuralHash, so new pictures can easily be added to the list. Once the tech is built into your iPhone, your MacBook and your Apple Watch, and can scan billions of photos a day, there will be pressure to use it for other purposes. The other part of NCMEC’s mission is missing children. Can Apple resist demands to help find runaways? Could Tim Cook possibly be so cold-hearted as to refuse at add Madeleine McCann to the watch list?

After that, your guess is as good as mine. Depending on where you are, you might find your photos scanned for dissidents, religious leaders or the FBI’s most wanted. It also reminds me of the Rasterfahndung in 1970s Germany – the dragnet search of all digital data in the country for clues to the Baader-Meinhof gang. Only now it can be done at scale, and not just for the most serious crimes either.

Finally, there’s adversarial machine learning. Neural networks are fairly easy to fool in that an adversary can tweak images so they’re misclassified. Expect to see pictures of cats (and of Tim Cook) that get flagged as abuse, and gangs finding ways to get real abuse past the system. Apple’s new tech may end up being a distributed person-search machine, rather than a sex-abuse prevention machine.

Such a technology requires public scrutiny, and as the possession of child sex abuse images is a strict-liability offence, academics cannot work with them. While the crooks will dig out NeuralMatch from their devices and play with it, we cannot. It is possible in theory for Apple to get NeuralMatch to ignore faces; for example, it could blur all the faces in the training data, as Google does for photos in Street View. But they haven’t claimed they did that, and if they did, how could we check? Apple should therefore publish full details of NeuralMatch plus a set of NeuralHash values trained on a public dataset with which we can legally work. It also needs to explain how the system it deploys was tuned and tested; and how dragnet searches of people’s photo libraries will be restricted to those conducted by court order so that they are proportionate, necessary and in accordance with the law. If that cannot be done, the technology must be abandoned.

Security engineering and machine learning

Last week I gave my first lecture in Edinburgh since becoming a professor there in February. It was also the first talk I’ve given in person to a live audience since February 2020.

My topic was the interaction between security engineering and machine learning. Many of the things that go wrong with machine-learning systems were already familiar in principle, as we’ve been using Bayesian techniques in spam filters and fraud engines for almost twenty years. Indeed, I warned about the risks of not being able to explain and justify the decisions of neural networks in the second edition of my book, back in 2008.

However the deep neural network (DNN) revolution since 2012 has drawn in hundreds of thousands of engineers, most of them without this background. Many fielded systems are extremely easy to break, often using tricks that have been around for years. What’s more, new attacks specific to DNNs – adversarial samples – have been found to exist for pretty well all models. They’re easy to find, and often transferable from one model to another.

I describe a number of new attacks and defences that we’ve discovered in the past three years, including the Taboo Trap, sponge attacks, data ordering attacks and markpainting. I argue that we will usually have to think of defences at the system level, rather than at the level of individual components; and that situational awareness is likely to play an important role.

Here now is the video of my talk.

A new way to detect ‘deepfake’ picture editing

Common graphics software now offers powerful tools for inpainting – using machine-learning models to reconstruct missing pieces of an image. They are widely used for picture editing and retouching, but like many sophisticated tools they can also be abused. They can remove someone from a picture of a crime scene, or remove a watermark from a stock photo. Could we make such abuses more difficult?

We introduce Markpainting, which uses adversarial machine-learning techniques to fool the inpainter into making its edits evident to the naked eye. An image owner can modify their image in subtle ways which are not themselves very visible, but will sabotage any attempt to inpaint it by adding visible information determined in advance by the markpainter.

One application is tamper-resistant marks. For example, a photo agency that makes stock photos available on its website with copyright watermarks can markpaint them in such a way that anyone using common editing software to remove a watermark will fail; the copyright mark will be markpainted right back. So watermarks can be made a lot more robust.

In the fight against fake news, markpainting news photos would mean that anyone trying to manipulate them would risk visible artefacts. So bad actors would have to check and retouch photos manually, rather than trying use inpainting tools to automate forgery at scale.

This paper has been accepted at ICML.

Robots, manners and stress

Humans and other animals have evolved to be aware of whether we’re under threat. When we’re on safe territory with family and friends we relax, but when we sense that a rival or a predator might be nearby, our fight-or-flight response kicks in. Situational awareness is vital, as it’s just too stressful to be alert all the time.

We’ve started to realise that this is likely to be just as important in many machine-learning applications. Take as an example machine vision in an automatic driver assistance system, whose goal is automatic lane keeping and automatic emergency braking. Such systems use deep neural networks, as they perform way better than the alternatives; but they can be easily fooled by adversarial examples. Should we worry? Sure, a bad person might cause a car crash by projecting a misleading image on a motorway bridge – but they could as easily steal some traffic cones from the road works. Nobody sits up at night worrying about that. But the car industry does actually detune vision systems from fear of deceptive attacks!

We therefore started a thread of research aimed at helping machine-learning systems detect whether they’re under attack. Our first idea was the Taboo Trap. You raise your kids to observe social taboos – to behave well and speak properly – and yet once you send them to school they suddenly know words that would make your granny blush. The taboo violation shows they’ve been exposed to ‘adversarial inputs’, as an ML engineer would call them. So we worked out how to train a neural network to avoid certain taboo values, both of outputs (forbidden utterances) and intermediate activations (forbidden thoughts). The taboos can be changed every time you retrain the network, giving the equivalent of a cryptographic key. Thus even though adversarial samples will always exist, you can make them harder to find; an attacker can’t just find one that works against one model of car and use it against every other model. You can take a view, based on risk, of how many different keys you need.

We then showed how you can also attack the availability of neural networks using sponge examples – inputs designed to soak up as much energy, and waste as much time, as possible. An alarm can be simpler to build in this case: just monitor how long your classifier takes to run.

Are there broader lessons? We suspect so. As robots develop situational awareness, like humans, and react to real or potential attacks by falling back to a more cautious mode of operation, a hostile environment will cause the equivalent of stress. Sometimes this will be deliberate; one can imagine constant low-level engagement between drones at tense national borders, just as countries currently probe each others’ air defences. But much of the time it may well be a by-product of poor automation design coupled with companies hustling aggressively for consumers’ attention.

This suggests a missing factor in machine-learning research: manners. We’ve evolved manners to signal to others that our intent is not hostile, and to negotiate the many little transactions that in a hostile environment might lead to a tussle for dominance. Yet these are hard for robots. Food-delivery robots can become unpopular for obstructing and harassing other pavement users; and one of the show-stoppers for automated driving is the difficulty that self-driving cars have in crossing traffic, or otherwise negotiating precedence with other road users. And even in the military, manners have a role – from the chivalry codes of medieval knights to the more modern protocols whereby warships and warplanes warn other craft before opening fire. If we let loose swarms of killer drones with no manners, conflict will be more likely.

Our paper Situational Awareness and Machine Learning – Robots, Manners and Stress was invited as a keynote for two co-located events: IEEE CogSIMA and the NATO STO SCI-341 Research Symposium on Situation awareness of Swarms and Autonomous systems. We got so many conflicting demands from the IEEE that we gave up on making a video of the talk for them, and our paper was pulled from their proceedings. However we decided to put the paper online for the benefit of the NATO folks, who were blameless in this matter.

Data ordering attacks

Most deep neural networks are trained by stochastic gradient descent. Now “stochastic” is a fancy Greek word for “random”; it means that the training data are fed into the model in random order.

So what happens if the bad guys can cause the order to be not random? You guessed it – all bets are off. Suppose for example a company or a country wanted to have a credit-scoring system that’s secretly sexist, but still be able to pretend that its training was actually fair. Well, they could assemble a set of financial data that was representative of the whole population, but start the model’s training on ten rich men and ten poor women drawn from that set – then let initialisation bias do the rest of the work.

Does this generalise? Indeed it does. Previously, people had assumed that in order to poison a model or introduce backdoors, you needed to add adversarial samples to the training data. Our latest paper shows that’s not necessary at all. If an adversary can manipulate the order in which batches of training data are presented to the model, they can undermine both its integrity (by poisoning it) and its availability (by causing training to be less effective, or take longer). This is quite general across models that use stochastic gradient descent.

This work helps remind us that computer systems with DNN components are still computer systems, and vulnerable to a wide range of well-known attacks. A lesson that cryptographers have learned repeatedly in the past is that if you rely on random numbers, they had better actually be random (remember preplay attacks) and you’d better not let an adversary anywhere near the pipeline that generates them (remember injection attacks). It’s time for the machine-learning community to carefully examine their assumptions about randomness.

Three Paper Thursday: Subverting Neural Networks via Adversarial Reprogramming

This is a guest post by Alex Shepherd.

Five years after Szegedy et al. demonstrated the capacity for neural networks to be fooled by crafted inputs containing adversarial perturbations, Elsayed et al. introduced adversarial reprogramming as a novel attack class for adversarial machine learning. Their findings demonstrated the capacity for neural networks to be reprogrammed to perform tasks outside of their original scope via crafted adversarial inputs, creating a new field of inquiry for the fields of AI and cybersecurity.

Their discovery raised important questions regarding the topic of trustworthy AI, such as what the unintended limits of functionality are in machine learning models and whether the complexity of their architectures can be advantageous to an attacker. For this Three Paper Thursday, we explore the three most eminent papers concerning this emerging threat in the field of adversarial machine learning.

Adversarial Reprogramming of Neural Networks, Gamaleldin F. Elsayed, Ian Goodfellow, and Jascha Sohl-Dickstein, International Conference on Learning Representations, 2018.

In their seminal paper, Elsayed et al. demonstrated their proof-of-concept for adversarial reprogramming by successfully repurposing six pre-trained ImageNet classifiers to perform three alternate tasks via crafted inputs containing adversarial programs. Their threat model considered an attacker with white-box access to the target models, whose objective was to subvert the models by repurposing them to perform tasks they were not originally intended to do. For the purposes of their hypothesis testing, adversarial tasks included counting squares and classifying MNIST digits and CIFAR-10 images.
Continue reading Three Paper Thursday: Subverting Neural Networks via Adversarial Reprogramming

Pushing the limits: acoustic side channels

How far can we go with acoustic snooping on data?

Seven years ago we showed that you could use a phone camera to measure the phone’s motion while typing and use that to recover PINs. Four years ago we showed that you could use interrupt timing to recover text entered using gesture typing. Last year we showed how a gaming app can steal your banking PIN by listening to the vibration of the screen as your finger taps it. In that attack we used the on-phone microphones, as they are conveniently located next to the screen and can hear the reverberations of the screen glass.

This year we wondered whether voice assistants can hear the same taps on a nearby phone as the on-phone microphones could. We knew that voice assistants could do acoustic snooping on nearby physical keyboards, but everyone had assumed that virtual keyboards were so quiet as to be invulnerable.

Almos Zarandy, Ilia Shumailov and I discovered that attacks are indeed possible. In Hey Alexa what did I just type? we show that when sitting up to half a meter away, a voice assistant can still hear the taps you make on your phone, even in presence of noise. Modern voice assistants have two to seven microphones, so they can do directional localisation, just as human ears do, but with greater sensitivity. We assess the risk and show that a lot more work is needed to understand the privacy implications of the always-on microphones that are increasingly infesting our work spaces and our homes.

Three Paper Thursday: Vēnī, Vīdī, Vote-y – Election Security

With the recent quadrennial instantiation of the US presidential election, discussions of election security have predictably resurged across much of the world. Indeed, news cycles in the US, UK, and EU abound with talking points surrounding the security of elections. In light of this context, we will use this week’s Three Paper Thursday to shed light on the technical challenges, solutions, and opportunities in designing secure election systems.

This post will focus on the technical security of election systems. That said, the topic of voter manipulation techniques such as disinformation campaigns, although out of scope here, is also an open area of research.

At first glance, voting may not seem like a challenging problem. If we are to consider a simple majority vote, surely a group of young schoolchildren could reach a consensus in minutes via hand-raising. Striving for more efficient vote tallying, though, perhaps we may opt to follow the IETF in consensus through humming. As we seek a solution that can scale to large numbers of voters, practical limitations will force us to select a multi-location, asynchronous process. Whether we choose in-person polling stations or mail-in voting, challenges quickly develop: how do we know a particular vote was counted, its contents kept secret, and the final tally correct?

National Academies of Sciences, Engineering, and Medicine (U.S.), Ed., Securing the vote: protecting American democracy, The National Academies Press (2018)

The first paper is particularly prominent due to its unified, no-nonsense, and thorough analysis. The report is specific to the United States, but its key themes apply generally. Written in response to accusations of international interference in the US 2016 presidential election, the National Academies provide 41 recommendations to strengthen the US election system.

These recommendations are extremely straightforward, and as such a reminder that adversaries most often penetrate large systems by targeting the “weakest link.” Among other things, the authors recommend creating standardized ballot data formats, regularly validating voter registration lists, evaluating the accessibility of ballot formats, ensuring access to absentee ballots, conducting appropriate audits, and providing adequate funding for elections.

It’s important to get the basics right. While there are many complex, stimulating proposals that utilize cutting-edge algorithms, cryptography, and distributed systems techniques to strengthen elections, many of these proposals are moot if the basic logistics are mishandled.

Some of these low-tech recommendations are, to the surprise of many passionate technologists, quite common among election security specialists. For example, requiring a paper ballot trail and avoiding internet voting based on current technology is also cited in our next paper.

Matthew Bernhard et al., Public Evidence from Secret Ballots, arXiv:1707.08619 (2017)

Governance aside, the second paper offers a comprehensive survey of the key technical challenges in election security and common tools used to solve them. The paper motivates the difficulty of election systems by attesting that all actors involved in an election are mutually distrustful, meaningful election results require evidence, and voters require ballot secrecy.

Ballot secrecy is more than a nicety; it is key to a properly functioning election system. Implemented correctly, ballot secrecy prevents voter coercion. If a voter’s ballot is not secret, or indeed if there is any way a voter can post-facto prove the casting a certain vote, malicious actors may pressure the voter to provide proof that they voted as directed. This can be insidiously difficult to prevent if not considered thoroughly.

Bernhard et al. discuss risk-limiting audits (RLAs) as an efficient yet powerful way to limit uncertainty in election results. By sampling and recounting a subset of votes, RLAs enable the use of statistical methods to increase confidence in a correct ballot count. Employed properly, RLAs can enable the high-probability validation of election tallies with effort inversely proportional to the expected margin. RLAs are now being used in real-world elections, and many RLA techniques exist in practice. 

Refreshingly, this paper establishes that blockchain-based voting is a bad idea. Blockchains inherently lack a central authority, so enforcing election rules would be a challenge. Furthermore, a computationally powerful adversary could control which votes get counted.

The paper also discusses high-level cryptographic tools that can be useful in elections. This leads us to our third and final paper.

Josh Benaloh, ElectionGuard Specification v0.95, Microsoft GitHub (2020)

Our final paper is slightly different from the others in this series; it’s a snapshot of a formal specification that is actively being developed, largely based on the author’s 1996 Yale doctoral thesis.

The specification describes ElectionGuard, a system being built by Microsoft to enable verifiable election results (disclaimer: the author of this post holds a Microsoft affiliation). It uses a combination of exponential ElGamal additively-homomorphic encryption, zero knowledge proofs, and Shamir’s secret sharing to conduct publicly-verifiable, secret-ballot elections.

When a voter casts a ballot, they are given a tracking code which can be used to verify the counting of the ballot’s votes via cryptographic proofs published with the final tally. Voters can achieve high confidence that their ballot represents a proper encryption of their desired votes by optionally spoiling an unlimited number of ballots triggering a decryption of the spoiled ballot at the time of voting. Encrypted ballots are homomorphically tallied in encrypted form by the election authorities, and the number of authorities that participate in tallying must meet the threshold set for the election to protect against malicious authorities.

The specification does not require that the system be used for exclusively internet-based or polling station-based elections; rather it is a framework for users to consume as they wish. Indeed, one of the draws to ElectionGuard is that it does not mandate a specific UI, ballot marking device, or even API. This flexibility allows election authorities to leverage the system in the manner that best fits their jurisdiction. The open source implementation can be found on GitHub.

There are many pieces of voting software available, but ElectionGuard is the new kid on the block that addresses many of the concerns raised in our earlier papers.

Key Themes

Designing secure election systems is difficult.

Often, election systems fall short on the basics; improper voting lists, postage issues, and poorly formatted ballots can disrupt elections as much as some adversaries. Ensuring that the foundational components of an election are handled well currently involves seemingly mundane but important things such as paper ballot trails, chains of custody, and voter ID verification.

High-tech election proposals are not new; indeed key insights into the use of cryptographic techniques in elections were being discussed in the academic literature well over two decades ago. That said, in recent years there has been an ostensibly increased investment in implementing cryptographic election systems, and although there remain many problems to be solved the future in this area looks promising.

Three Paper Thursday: Attacking Machine Vision Models In Real Life

This is a guest post by Alex Shepherd.

There is a growing body of research literature concerning the potential threat of physical-world adversarial attacks against machine-vision models. By applying adversarial perturbations to physical objects, machine-vision models may be vulnerable to images containing these perturbed objects, resulting in an increased risk of misclassification. The potential impacts could be significant and have been identified as risk areas for autonomous vehicles and military UAVs.

For this Three Paper Thursday, we examine the following papers exploring the potential threat of physical-world adversarial attacks, with a focus on the impact for autonomous vehicles.

Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world, arXiv:1607.02533 (2016)

In this seminal paper, Kurakin et al. report their findings of an experiment conducted using adversarial images taken from a phone camera as input for a pre-trained ImageNet Inceptionv3 image classification model. Methodology was based on a white-box threat model, with adversarial images crafted from the ImageNet validation dataset using the Inceptionv3 model.
Continue reading Three Paper Thursday: Attacking Machine Vision Models In Real Life

SRI and Cambridge release CHERI software stack for Arm Morello

For the last ten years, SRI International and the University of Cambridge have been working to develop CHERI (Capability Hardware Enhanced RISC Instructions), a DARPA-sponsored processor architecture security technology implementing efficient fine-grained memory protection and scalable software compartmentalization. You can learn more about CHERI in our Introduction to CHERI technical report, which describes the architectural, microarchitectural, formal modelling, and software approaches we have created.

For the last six of those years, we have been collaborating closely with Arm to create an adaptation of CHERI to the ARMv8-A architecture, which is slated to appear in Arm’s prototype Morello processor, System-on-Chip (SoC), and board in Q1 2022. Richard Grisenthwaite, Arm’s Principal Architect, announced this joint work at the UKRI Digital Security by Design (DSbD) workshop in September 2019. DSbD is a UKRI / Industrial Strategy Challenge Fund (ISCF) research programme contributing to the creation of the Morello board, and CHERI is the Digital Security by Design Technology that underlies the programme. Our collaboration with Arm has been an enormously exciting experience, involving daily engagement Arm’s architects, microarchitects, and software designers. This included hosting several members of Arm’s team at our lab in Cambridge over multiple years, as we brought together our long-term research on architectural and software security with their experience in industrial architecture, processor designs, and transition.

Today, Richard Grisenthwaite announced that Arm is releasing their first simulator for the Morello architecture, the Morello FVP (Fixed Virtual Platform), and also an open-source software stack that includes their adaptation of our CHERI Clang/LLVM to Morello and early work on Morello support for Android. These build on the Morello architecture specification, released in late September 2020. SRI and Cambridge are releasing a first developer preview release of the CHERI reference software stack ported to Morello – intended to show a rich integration of CHERI into a contemporary OS design, as well as demonstration applications. This stack includes CheriBSD, a BSD-licensed reference design and open-source applications adapted to CHERI including OpenSSH, nginx, and WebKit.

For this first developer preview release, we have focused on bringing CHERI C/C++ memory protection to Morello. Our CheriABI process environment, which allows the full UNIX userspace to run with fine-grained spatial memory safety, is fully functional on Morello. This work has been the recent subject of a report from the Microsoft Security Response Center (MSRC), Microsoft’s internal red team and security response organization, describing how CHERI has to potential to deterministically prevent over 2/3 of critical Microsoft software security vulnerabilities. CheriBSD/Morello brings that work over from our research CHERI-MIPS and CHERI-RISC-V platforms to Arm’s Morello. We demonstrated CheriBSD/Morello mitigating several memory-safety vulnerabilities in the EPSRC Digital Security by Design (DSbD) workshop yesterday, talking to 9 UK universities that have been funded to do research building on CHERI and Morello.

We have an aggressive planned quarterly release schedule through the end of 2021 when a full release will ship alongside the Morello board, adapting various CheriBSD security features to Morello:

DateReleaseKey features
October 2020Developer PreviewCheriABI pure-capability userspace implementing spatial memory safety.
December 2020Update 1Pure-capability kernel implementing spatial memory safety.
March 2021Update 2Userspace heap temporal memory safety based on Cornucopia (in collaboration with Microsoft Research).
June 2021Update 3Userspace software compartmentalization based on the CHERI co-process model.
October 2021Update 4Userspace software compartmentalization based on a run-time linker model.
Late 2021Full releaseAny updates required to operate well on the shipping Morello board.
CHERI software stack – working release schedule for 2020-2021

Getting started with CheriBSD/Morello is easy (if you have a tolerance for experimental architectural simulators, experimental operating systems, and experimental compilers!). Visit our CHERI Morello software web page to learn more about this work, and then our CheriBSD/Morello distribution page to download our build environment. You can automatically install Arm’s FVP, cross-develop in our docker-based SDK on macOS or Linux, and SSH into the simulated host to try things out.

CHERI is the work of a large research team at SRI International and the University of Cambridge, as well as numerous industrial collaborators at Arm, Google, Microsoft, and elsewhere. My co-investigators, Peter G. Neumann (SRI), Simon W. Moore (Cambridge), Peter Sewell (Cambridge), and I are immensely grateful for their contributions: CHERI would simply not have been possible without your collective effort – thank you! We are also grateful to our sponsors over an extended period, including DARPA, UKRI, Google, and Arm.