Robust Physical-World Attacks on Machine Learning Modules

Adv ML Website Carousel

Could graffiti convey a hidden message to your car? Or cause a robot to do something unexpected? Cars and robots, as well as other devices, are more frequently relying on images of their surroundings to make decisions. New research explores the possibility that malicious alterations to real world objects, like the road sign above, could cause these devices to “misread” the image and take a certain adverse action. The paper Robust Physical-World Attacks on Deep Learning Modules is by a research team spanning the University of Washington, including Ph.D. student Ivan Etimov, Tech Policy Lab postdoc Earlence Fernandes, and Co-Director Yoshi Kohno, along with Kevin Eykholt and Atul Prakash from the University of Michigan, Amir Rahmati from Stony Brook University, and from the University of California Berkeley Bo Li and Dawn Song.

To address this question, the researchers created an algorithm that could generate these alterations, a methodology to evaluate their effectiveness in fooling machine learning, and then applied both to the real world example of autonomous vehicles. They experimented to see whether physically altering an object, in this case a road sign, could cause the computer of an autonomous vehicle to classify it incorrectly. Autonomous vehicles learn to classify objects using machine learning, where the car’s computer “learns” what objects such as road signs, pedestrians, and other cars look like by being shown thousands of photos of each object. If you’re not familiar with machine learning, check out the Lab’s fun primer video “What is Machine Learning?” here. Current self driving car systems can include this type of camera sensor, as well as a variety of others such as lidar, radar, and GPS.

The researchers wanted to explore whether it’s possible to fool these machine learning “brains” by slightly altering images shown to the classifier, which identifies, in the case of autonomous vehicles, the different road signs seen by the car’s camera sensors. While previous research has focused on altering an image digitally and then feeding that digital image into a classifier, the research team wanted to see if it was possible to physically, rather than digitally, alter the content of the image to maliciously fool the classifier.

In order to generate ways to alter these road signs, the researchers applied their new algorithm that looked at what the trained classifier “knew” about road signs, and generated ways to alter the signs that would fool the classifier when used in the real world. The research focuses on two types of alterations generated by the algorithm:
• poster-printing attacks, where an attacker prints an actual-sized poster of a road sign that has subtle variations and pastes it over the real sign, and
• sticker attacks, where an attacker prints the generated sticker design and places it onto the existing road sign.

Poster Printing Attack                                                                    Sticker Attack

Following their proposed methodology, the researchers took photos of the signs from a range of physical conditions that mimic different positions under which a sensor might encounter the object, and then fed those images into a machine learning application, in this case a road sign classifier. When photos of the above stop signs taken from different angles and distances were fed into the researchers’ road sign classifier in lab testing, the classifier misread them as speed limit signs 100% of the time for the poster-printing attack, and 66% of the time for the sticker attack. Because these attacks mimic vandalism or street art, it can be difficult for a casual observer to identify the risk they could pose.

The researchers show that it is possible to generate real world alterations to objects that fool machine learning under a variety of conditions. They propose a new methodology for evaluating the effectiveness of these alterations under a range of diverse physical conditions that mimic those a sensor may encounter the object under in the real world. The researchers’ aim is to help improve the security of technology like autonomous vehicles in the future, by identifying security risks now. To read more, see the paper Robust Physical-World Attacks on Deep Learning Models as well as the FAQ .

Privacy in Online Dating

How do you manage your privacy in online dating? Chances are that if you use online dating or have considered using it, this is an issue you’ve given some thought. And you wouldn’t be alone, as privacy issues in online dating have appeared in the media—two summers ago, during the Rio Olympics, privacy in online dating made headlines when a Tinder user posted screenshots of Olympian’s profiles on social media, and a journalist collected identifying information about closeted gay Olympians through Grindr. In September, a journalist requested her personal data from Tinder and received 800+ pages including information about her Facebook and Instagram activity. And more recently, researchers have revealed security vulnerabilities in a number of online dating apps, including ways that users may be vulnerable due to sensitive information they disclose on the site.

These events and others show that individual users can’t control all privacy-related risks when using online dating. To understand how users reason about privacy risks they can potentially control through decision making, Lab Ph.D. student Camille Cobb and Lab Faculty Co-Director Yoshi Kohno studied online dating user’s perceptions about and actions governing their privacy in “How Public is My Private Life? Privacy in Online Dating.” The researchers surveyed 100 participants about how they handle their own and other users’ privacy; then, based on themes raised in survey responses, they conducted follow-up interviews and analyzed a sample of Tinder profiles.


Based on their survey of 100 online dating users, and interviews with 14 of those, the researchers found that when choosing profile content, looking people up, and taking screenshots of messages or profiles, users may face complex tradeoffs between preserving their own or others’ privacy and other goals. Users described balancing privacy considerations including the risk of feeling awkward, screenshots and data breaches, stalking, and their profile being seen by a friend or co-worker, with goals like getting successful matches, preserving information that may be sentimental if the match is successful, safety, and avoiding scams.

These tradeoffs are complex, and involve user’s privacy decisions beyond just the dating app. For example, a user concerned about privacy on social media like Facebook might change their name to something unusual or unique and hard to guess, but if a dating service pulls that name into a user’s profile, that unique name makes the user easier to find outside of the dating app. Beyond the name they used, users also experienced tradeoffs around the amount of information to include in their profile. Including more information could make users more easily searchable outside of the dating app, while not including any identifiable information could run the risk, in one user’s case, of being thought to be a bot.

Focusing on users’ concerns about “searchibility,” or the risk of being identified elsewhere online, the researchers analyzed 400 Tinder profiles. Using techniques readily available and fairly easy for any Tinder user to use, event without technical knowledge, the researchers were able to find 47% of the users. And having an account directly linked to another account, or mentioning a username for another account in the profile, increased the chance of being found to 80%. These results support concerns suggested in the survey; the researchers were able to find a larger portion of people with unique names, echoing a survey respondent’s concern that having a unique name would make her more identifiable.

Discussing the privacy considerations & tradeoffs that users described experiencing, and in light of their analysis of profiles’ searchability, the researchers suggest a number of avenues to explore that could help online dating users make decisions around privacy. These could include restricting the number of screenshots a user can take per day, allowing users to disallow remove matches, and, more broadly, implementing privacy awareness campaigns for users.

This paper was presented at the 26th International World-Wide Web Conference and is available here.

Tech Policy Lab Joins Partnership on Artificial Intelligence

The Tech Policy Lab is delighted to be joining the Partnership on AI to Benefit People and Society, a non-profit organization charged with exploring and developing best practices for AI. The Lab, which aims to position policymakers, broadly defined, to make wiser and more inclusive tech policy, joins a diverse range of voices from academia, industry and non-profit organizations committed to collaboration and open dialogue on the opportunities and rising challenges around AI.

PAI Logo_Large_2

The Lab has worked to advance AI in the public interest since our inception, through conferences, workshops, and research, among other initiatives.  In 2015, we organized the fourth annual robotics law and policy conference, WeRobot. And in 2016, we co-organized the Obama White House’s inaugural public workshop on AI, focusing on legal and governance implications of AI.  Our research focuses on the policy implications of AI and includes studying AI-connected devices in the home.

We are planning many more research initiatives around AI, including AI-assisted decision-making, AI and cybersecurity, and AI and diversity. We will bring to our AI research our commitment to the inclusion of diverse perspectives in tech policy research and outcomes, including our Diverse Voices method (made available earlier this year through our How-To Guide) which engages diverse panels of “experiential” experts in short, targeted conversations around a technology to improve inclusivity in tech policy outcomes.

The Partnership on AI will be a great network and resource as we undertake this work. We look forward to collaborating with a diverse group of stakeholders from industry, academia, and policy around the Partnership on AI’s goals: to develop and share best practices, advance public understanding of AI, create a diverse network of experts around AI, and examine AI’s impact on people and society.

About the Tech Policy Lab

The Tech Policy Lab is a unique, interdisciplinary research unit at the University of Washington. The Lab’s mission is to position policymakers, broadly defined, to make wiser and more inclusive tech policy.  Situated within a globally renowned research university, the Tech Policy Lab is committed to advancing artificial intelligence in the public interest through research, analysis, and education and outreach. To learn more about the Lab’s cutting edge research, thought leadership, and education initiatives, go to

About the Partnership on AI

The Partnership on AI to Benefit People and Society (Partnership on AI) is a not-for-profit organization, founded by Amazon, Apple, Google/DeepMind, Facebook, IBM and Microsoft.  Our goals are to study and formulate best practices on the development, testing, and fielding of AI technologies, advancing the public’s understanding of AI, to serve as an open platform for discussion and engagement about AI and its influences on people and society and identify and foster aspirational efforts in AI for socially beneficial purposes. We actively designed the Partnership on AI to bring together a diverse range of voices from for-profit and non-profit, all of whom share our belief in the tenets and are committed to collaboration and open dialogue on the many opportunities and rising challenges around AI. For the full list of founding members and partners, go to

Exploring ADINT: Using Ad Targeting for Surveillance on a Budget

New research by former CSE Ph.D. student Paul Vines, Lab Faculty Associate Franzi Roesner, and Faculty Co-Director Yoshi Kohno demonstrates how targeted advertising can be used for personal surveillance.

From “Exploring ADINT: Using Ad Targeting for Surveillance on a Budget – or – How Alice Can Buy Ads to Track Bob

The online advertising ecosystem is built upon the ability of advertising networks to know properties about users (e.g., their interests or physical locations) and deliver targeted ads based on those properties. Much of the privacy debate around online advertising has focused on the harvesting of these properties by the advertising networks. In this work, we explore the following question: can third-parties use the purchasing of ads to extract private information about individuals? We find that the answer is yes. For example, in a case study with an archetypal advertising network, we find that — for $1000 USD — we can track the location of individuals who are using apps served by that advertising network, as well as infer whether they are using potentially sensitive applications (e.g., certain religious or sexuality-related apps). We also conduct a broad survey of other ad networks and assess their risks to similar attacks. We then step back and explore the implications of our findings.

The Tech Policy Lab plans to work with the ADINT research team to explore the policy implications of this research, examining potential recommendations for issues raised by this new form of personal surveillance.

More information can be found on the team’s website, and the UW News and UW CSE releases. The paper will be presented at ACM’s Workshop on Privacy in the Electronic Society later this month and can be found here.

Securing Augmented Reality Output

A year ago, Pokemon Go became immensely popular as players explored their surroundings for Pokemon in the smartphone-based augmented reality (AR) app. This hyper-popular game, which barely scratched the surface of AR’s potential, led to increased interest in the technology. The AR industry is expected to grow to $100 billion by 2020, and with increasing interest in AR automotive windshields and head-mounted displays (HMDs), we could soon be able to experience immersive AR environments like the one depicted by designer and film-maker Keiichi Matsuda in Hyper Reality.

Hyper Reality Screenshot
But what would happen if a pop-up ad covers your game, causing you to lose? Or if, while driving, an AR object obscures a pedestrian?

These are the types of situations researchers consider in a new paper, Securing Augmented Reality Output. In the paper, Lab student Kiron Lebeck, along with CSE undergraduate Kimberly Ruth, Lab Affiliate Faculty Franzi Roesner, and Lab Co-Director Yoshi Kohno address how to defend against buggy or malicious AR software that may unintentionally or inadvertently augment a user’s view of the world in undesirable or harmful ways. They ask, how can we enable the operating system of an AR platform to play a role in mitigating these kinds of risks? To address this issue, the team designed Arya, an AR platform that controls output through a designated policy framework, drawing policy conditions from a range of sources including the Microsoft HoloLens development guidelines and the National Highway Traffic and Safety Administration (NHTSA)’s driver distraction guidelines.

Arya Driving Scenario
By identifying specific “if-then” policy statements, this policy framework allows the Arya platform to apply a specific mechanism, or action, to virtual objects that violate a condition. In a simulated driving experience, for example, Arya makes transparent pop-up ads and notifications that could distract the driver by applying a specified action, in this case transparency, to objects that violate the specific policies:
• Don’t obscure pedestrians,
• Only allow ads to appear on billboards, and
• Don’t distract the user while driving.

By implementing Arya in a prototype AR operating system, the team was able to prevent undesirable behavior in case studies of three environments, including a simulated driving scenario. Additionally, performance overhead of policy enforcement is acceptable even in the un-optimized prototype. The team, among the first to raise AR output security issues, demonstrated the feasibility of implementing a policy framework to address AR output security risks, while also surfacing lessons and directions for future efforts in the AR security space.

To read more, see Securing Augmented Reality Output.