Michael Backes

Wai Man Si, Michael Backes, Jeremy Blackburn et al.

Chatbots are used in many applications, e.g., automated agents, smart home assistants, interactive characters in online games, etc. Therefore, it is crucial to ensure they do not behave in undesired manners, providing offensive or toxic responses to users. This is not a trivial task as state-of-the-art chatbot models are trained on large, public datasets openly collected from the Internet. This paper presents a first-of-its-kind, large-scale measurement of toxicity in chatbots. We show that publicly available chatbots are prone to providing toxic responses when fed toxic queries. Even more worryingly, some non-toxic queries can trigger toxic responses too. We then set out to design and experiment with an attack, ToxicBuddy, which relies on fine-tuning GPT-2 to generate non-toxic queries that make chatbots respond in a toxic manner. Our extensive experimental evaluation demonstrates that our attack is effective against public chatbot models and outperforms manually-crafted malicious queries proposed by previous work. We also evaluate three defense mechanisms against ToxicBuddy, showing that they either reduce the attack performance at the cost of affecting the chatbot's utility or are only effective at mitigating a portion of the attack. This highlights the need for more research from the computer security and online safety communities to ensure that chatbot models do not hurt their users. Overall, we are confident that ToxicBuddy can be used as an auditing tool and that our work will pave the way toward designing more effective defenses for chatbot safety.

Boyang Zhang, Yicong Tan, Yun Shen et al.

Recently, autonomous agents built on large language models (LLMs) have experienced significant development and are being deployed in real-world applications. These agents can extend the base LLM's capabilities in multiple ways. For example, a well-built agent using GPT-3.5-Turbo as its core can outperform the more advanced GPT-4 model by leveraging external components. More importantly, the usage of tools enables these systems to perform actions in the real world, moving from merely generating text to actively interacting with their environment. Given the agents' practical applications and their ability to execute consequential actions, it is crucial to assess potential vulnerabilities. Such autonomous systems can cause more severe damage than a standalone language model if compromised. While some existing research has explored harmful actions by LLM agents, our study approaches the vulnerability from a different perspective. We introduce a new type of attack that causes malfunctions by misleading the agent into executing repetitive or irrelevant actions. We conduct comprehensive evaluations using various attack methods, surfaces, and properties to pinpoint areas of susceptibility. Our experiments reveal that these attacks can induce failure rates exceeding 80\% in multiple scenarios. Through attacks on implemented and deployable agents in multi-agent scenarios, we accentuate the realistic risks associated with these vulnerabilities. To mitigate such attacks, we propose self-examination detection methods. However, our findings indicate these attacks are difficult to detect effectively using LLMs alone, highlighting the substantial risks associated with this vulnerability.

Yuan Xin, Zheng Li, Ning Yu et al.

Despite being prevalent in the general field of Natural Language Processing (NLP), pre-trained language models inherently carry privacy and copyright concerns due to their nature of training on large-scale web-scraped data. In this paper, we pioneer a systematic exploration of such risks associated with pre-trained language encoders, specifically focusing on the membership leakage of pre-training data exposed through downstream models adapted from pre-trained language encoders-an aspect largely overlooked in existing literature. Our study encompasses comprehensive experiments across four types of pre-trained encoder architectures, three representative downstream tasks, and five benchmark datasets. Intriguingly, our evaluations reveal, for the first time, the existence of membership leakage even when only the black-box output of the downstream model is exposed, highlighting a privacy risk far greater than previously assumed. Alongside, we present in-depth analysis and insights toward guiding future researchers and practitioners in addressing the privacy considerations in developing pre-trained language models.

Min Chen, Zhikun Zhang, Tianhao Wang et al.

Machine unlearning is a process of removing the impact of some training data from the machine learning (ML) models upon receiving removal requests. While straightforward and legitimate, retraining the ML model from scratch incurs a high computational overhead. To address this issue, a number of approximate algorithms have been proposed in the domain of image and text data, among which SISA is the state-of-the-art solution. It randomly partitions the training set into multiple shards and trains a constituent model for each shard. However, directly applying SISA to the graph data can severely damage the graph structural information, and thereby the resulting ML model utility. In this paper, we propose GraphEraser, a novel machine unlearning framework tailored to graph data. Its contributions include two novel graph partition algorithms and a learning-based aggregation method. We conduct extensive experiments on five real-world graph datasets to illustrate the unlearning efficiency and model utility of GraphEraser. It achieves 2.06$\times$ (small dataset) to 35.94$\times$ (large dataset) unlearning time improvement. On the other hand, GraphEraser achieves up to $62.5\%$ higher F1 score and our proposed learning-based aggregation method achieves up to $112\%$ higher F1 score.\footnote{Our code is available at \url{https://github.com/MinChen00/Graph-Unlearning}.}

Saba Eskandarian, Jonathan Cogan, Sawyer Birnbaum et al.

Users regularly enter sensitive data, such as passwords, credit card numbers, or tax information, into the browser window. While modern browsers provide powerful client-side privacy measures to protect this data, none of these defenses prevent a browser compromised by malware from stealing it. In this work, we present Fidelius, a new architecture that uses trusted hardware enclaves integrated into the browser to enable protection of user secrets during web browsing sessions, even if the entire underlying browser and OS are fully controlled by a malicious attacker. Fidelius solves many challenges involved in providing protection for browsers in a fully malicious environment, offering support for integrity and privacy for form data, JavaScript execution, XMLHttpRequests, and protected web storage, while minimizing the TCB. Moreover, interactions between the enclave and the browser, the keyboard, and the display all require new protocols, each with their own security considerations. Finally, Fidelius takes into account UI considerations to ensure a consistent and simple interface for both developers and users. As part of this project, we develop the first open source system that provides a trusted path from input and output peripherals to a hardware enclave with no reliance on additional hypervisor security assumptions. These components may be of independent interest and useful to future projects. We implement and evaluate Fidelius to measure its performance overhead, finding that Fidelius imposes acceptable overhead on page load and user interaction for secured pages and has no impact on pages and page components that do not use its enhanced security features.

Junjie Chu, Yugeng Liu, Ziqing Yang et al.

Jailbreak attacks aim to bypass the LLMs' safeguards. While researchers have proposed different jailbreak attacks in depth, they have done so in isolation -- either with unaligned settings or comparing a limited range of methods. To fill this gap, we present a large-scale evaluation of various jailbreak attacks. We collect 17 representative jailbreak attacks, summarize their features, and establish a novel jailbreak attack taxonomy. Then we conduct comprehensive measurement and ablation studies across nine aligned LLMs on 160 forbidden questions from 16 violation categories. Also, we test jailbreak attacks under eight advanced defenses. Based on our taxonomy and experiments, we identify some important patterns, such as heuristic-based attacks could achieve high attack success rates but are easy to mitigate by defenses, causing low practicality. Our study offers valuable insights for future research on jailbreak attacks and defenses. We hope our work could help the community avoid incremental work and serve as an effective benchmark tool for practitioners.

Mingjie Li, Wai Man Si, Michael Backes et al.

As advancements in large language models (LLMs) continue and the demand for personalized models increases, parameter-efficient fine-tuning (PEFT) methods (e.g., LoRA) will become essential due to their efficiency in reducing computation costs. However, recent studies have raised alarming concerns that LoRA fine-tuning could potentially compromise the safety alignment in LLMs, posing significant risks for the model owner. In this paper, we first investigate the underlying mechanism by analyzing the changes in safety alignment related features before and after fine-tuning. Then, we propose a fixed safety module calculated by safety data and a task-specific initialization for trainable parameters in low-rank adaptations, termed Safety-alignment preserved Low-Rank Adaptation (SaLoRA). Unlike previous LoRA methods and their variants, SaLoRA enables targeted modifications to LLMs without disrupting their original alignments. Our experiments show that SaLoRA outperforms various adapters-based approaches across various evaluation metrics in different fine-tuning tasks.

Xinyue Shen, Yixin Wu, Yiting Qu et al.

Large Language Models (LLMs) have raised increasing concerns about their misuse in generating hate speech. Among all the efforts to address this issue, hate speech detectors play a crucial role. However, the effectiveness of different detectors against LLM-generated hate speech remains largely unknown. In this paper, we propose HateBench, a framework for benchmarking hate speech detectors on LLM-generated hate speech. We first construct a hate speech dataset of 7,838 samples generated by six widely-used LLMs covering 34 identity groups, with meticulous annotations by three labelers. We then assess the effectiveness of eight representative hate speech detectors on the LLM-generated dataset. Our results show that while detectors are generally effective in identifying LLM-generated hate speech, their performance degrades with newer versions of LLMs. We also reveal the potential of LLM-driven hate campaigns, a new threat that LLMs bring to the field of hate speech detection. By leveraging advanced techniques like adversarial attacks and model stealing attacks, the adversary can intentionally evade the detector and automate hate campaigns online. The most potent adversarial attack achieves an attack success rate of 0.966, and its attack efficiency can be further improved by $13-21\times$ through model stealing attacks with acceptable attack performance. We hope our study can serve as a call to action for the research community and platform moderators to fortify defenses against these emerging threats.

Wai Man Si, Mingjie Li, Michael Backes et al.

Recent reasoning large language models (LLMs), such as OpenAI o1 and DeepSeek-R1, exhibit strong performance on complex tasks through test-time inference scaling. However, prior studies have shown that these models often incur significant computational costs due to excessive reasoning, such as frequent switching between reasoning trajectories (e.g., underthinking) or redundant reasoning on simple questions (e.g., overthinking). In this work, we expose a novel threat: adversarial inputs can be crafted to exploit excessive reasoning behaviors and substantially increase computational overhead without compromising model utility. Therefore, we propose a novel loss framework consisting of three components: (1) Priority Cross-Entropy Loss, a modification of the standard cross-entropy objective that emphasizes key tokens by leveraging the autoregressive nature of LMs; (2) Excessive Reasoning Loss, which encourages the model to initiate additional reasoning paths during inference; and (3) Delayed Termination Loss, which is designed to extend the reasoning process and defer the generation of final outputs. We optimize and evaluate our attack for the GSM8K and ORCA datasets on DeepSeek-R1-Distill-LLaMA and DeepSeek-R1-Distill-Qwen. Empirical results demonstrate a 3x to 9x increase in reasoning length with comparable utility performance. Furthermore, our crafted adversarial inputs exhibit transferability, inducing computational overhead in o3-mini, o1-mini, DeepSeek-R1, and QWQ models.

Junjie Chu, Mingjie Li, Ziqing Yang et al.

Accurately determining whether a jailbreak attempt has succeeded is a fundamental yet unresolved challenge. Existing evaluation methods rely on misaligned proxy indicators or naive holistic judgments. They frequently misinterpret model responses, leading to inconsistent and subjective assessments that misalign with human perception. To address this gap, we introduce JADES (Jailbreak Assessment via Decompositional Scoring), a universal jailbreak evaluation framework. Its key mechanism is to automatically decompose an input harmful question into a set of weighted sub-questions, score each sub-answer, and weight-aggregate the sub-scores into a final decision. JADES also incorporates an optional fact-checking module to strengthen the detection of hallucinations in jailbreak responses. We validate JADES on JailbreakQR, a newly introduced benchmark proposed in this work, consisting of 400 pairs of jailbreak prompts and responses, each meticulously annotated by humans. In a binary setting (success/failure), JADES achieves 98.5% agreement with human evaluators, outperforming strong baselines by over 9%. Re-evaluating five popular attacks on four LLMs reveals substantial overestimation (e.g., LAA's attack success rate on GPT-3.5-Turbo drops from 93% to 69%). Our results show that JADES could deliver accurate, consistent, and interpretable evaluations, providing a reliable basis for measuring future jailbreak attacks.

Xinyu Zhang, Yixin Wu, Boyang Zhang et al.

Images shared on social media often expose geographic cues. While early geolocation methods required expert effort and lacked generalization, the rise of Large Vision Language Models (LVLMs) now enables accurate geolocation even for ordinary users. However, existing approaches are not optimized for this task. To explore the full potential and associated privacy risks, we present Geo-Detective, an agent that mimics human reasoning and tool use for image geolocation inference. It follows a procedure with four steps that adaptively selects strategies based on image difficulty and is equipped with specialized tools such as visual reverse search, which emulates how humans gather external geographic clues. Experimental results show that GEO-Detective outperforms baseline large vision language models (LVLMs) overall, particularly on images lacking visible geographic features. In country level geolocation tasks, it achieves an improvement of over 11.1% compared to baseline LLMs, and even at finer grained levels, it still provides around a 5.2% performance gain. Meanwhile, when equipped with external clues, GEO-Detective becomes more likely to produce accurate predictions, reducing the "unknown" prediction rate by more than 50.6%. We further explore multiple defense strategies and find that Geo-Detective exhibits stronger robustness, highlighting the need for more effective privacy safeguards.

Yixin Wu, Rui Wen, Chi Cui et al.

Inference attacks have been widely studied and offer a systematic risk assessment of ML services; however, their implementation and the attack parameters for optimal estimation are challenging for non-experts. The emergence of advanced large language models presents a promising yet largely unexplored opportunity to develop autonomous agents as inference attack experts, helping address this challenge. In this paper, we propose AttackPilot, an autonomous agent capable of independently conducting inference attacks without human intervention. We evaluate it on 20 target services. The evaluation shows that our agent, using GPT-4o, achieves a 100.0% task completion rate and near-expert attack performance, with an average token cost of only $0.627 per run. The agent can also be powered by many other representative LLMs and can adaptively optimize its strategy under service constraints. We further perform trace analysis, demonstrating that design choices, such as a multi-agent framework and task-specific action spaces, effectively mitigate errors such as bad plans, inability to follow instructions, task context loss, and hallucinations. We anticipate that such agents could empower non-expert ML service providers, auditors, or regulators to systematically assess the risks of ML services without requiring deep domain expertise.

Yiting Qu, Xinyue Shen, Yixin Wu et al.

With the advent of text-to-image models and concerns about their misuse, developers are increasingly relying on image safety classifiers to moderate their generated unsafe images. Yet, the performance of current image safety classifiers remains unknown for both real-world and AI-generated images. In this work, we propose UnsafeBench, a benchmarking framework that evaluates the effectiveness and robustness of image safety classifiers, with a particular focus on the impact of AI-generated images on their performance. First, we curate a large dataset of 10K real-world and AI-generated images that are annotated as safe or unsafe based on a set of 11 unsafe categories of images (sexual, violent, hateful, etc.). Then, we evaluate the effectiveness and robustness of five popular image safety classifiers, as well as three classifiers that are powered by general-purpose visual language models. Our assessment indicates that existing image safety classifiers are not comprehensive and effective enough to mitigate the multifaceted problem of unsafe images. Also, there exists a distribution shift between real-world and AI-generated images in image qualities, styles, and layouts, leading to degraded effectiveness and robustness. Motivated by these findings, we build a comprehensive image moderation tool called PerspectiveVision, which improves the effectiveness and robustness of existing classifiers, especially on AI-generated images. UnsafeBench and PerspectiveVision can aid the research community in better understanding the landscape of image safety classification in the era of generative AI.

Carmela Troncoso, Mathias Payer, Jean-Pierre Hubaux et al.

This document describes and analyzes a system for secure and privacy-preserving proximity tracing at large scale. This system, referred to as DP3T, provides a technological foundation to help slow the spread of SARS-CoV-2 by simplifying and accelerating the process of notifying people who might have been exposed to the virus so that they can take appropriate measures to break its transmission chain. The system aims to minimise privacy and security risks for individuals and communities and guarantee the highest level of data protection. The goal of our proximity tracing system is to determine who has been in close physical proximity to a COVID-19 positive person and thus exposed to the virus, without revealing the contact's identity or where the contact occurred. To achieve this goal, users run a smartphone app that continually broadcasts an ephemeral, pseudo-random ID representing the user's phone and also records the pseudo-random IDs observed from smartphones in close proximity. When a patient is diagnosed with COVID-19, she can upload pseudo-random IDs previously broadcast from her phone to a central server. Prior to the upload, all data remains exclusively on the user's phone. Other users' apps can use data from the server to locally estimate whether the device's owner was exposed to the virus through close-range physical proximity to a COVID-19 positive person who has uploaded their data. In case the app detects a high risk, it will inform the user.

Ilkan Esiyok, Lucjan Hanzlik, Robert Kuennemann et al.

Astroturfing, i.e., the fabrication of public discourse by private or state-controlled sponsors via the creation of fake online accounts, has become incredibly widespread in recent years. It gives a disproportionally strong voice to wealthy and technology-savvy actors, permits targeted attacks on public forums and could in the long run harm the trust users have in the internet as a communication platform. Countering these efforts without deanonymising the participants has not yet proven effective; however, we can raise the cost of astroturfing. Following the principle `one person, one voice', we introduce Trollthrottle, a protocol that limits the number of comments a single person can post on participating websites. Using direct anonymous attestation and a public ledger, the user is free to choose any nickname, but the number of comments is aggregated over all posts on all websites, no matter which nickname was used. We demonstrate the deployability of Trollthrottle by retrofitting it to the popular news aggregator website Reddit and by evaluating the cost of deployment for the scenario of a national newspaper (168k comments per day), an international newspaper (268k c/d) and Reddit itself (4.9M c/d).

Yang Zhang, Mathias Humbert, Tahleen Rahman et al.

Hashtag has emerged as a widely used concept of popular culture and campaigns, but its implications on people's privacy have not been investigated so far. In this paper, we present the first systematic analysis of privacy issues induced by hashtags. We concentrate in particular on location, which is recognized as one of the key privacy concerns in the Internet era. By relying on a random forest model, we show that we can infer a user's precise location from hashtags with accuracy of 70\% to 76\%, depending on the city. To remedy this situation, we introduce a system called Tagvisor that systematically suggests alternative hashtags if the user-selected ones constitute a threat to location privacy. Tagvisor realizes this by means of three conceptually different obfuscation techniques and a semantics-based metric for measuring the consequent utility loss. Our findings show that obfuscating as little as two hashtags already provides a near-optimal trade-off between privacy and utility in our dataset. This in particular renders Tagvisor highly time-efficient, and thus, practical in real-world settings.

Sanam Ghorbani Lyastani, Michael Schilling, Sascha Fahl et al.

Despite their well-known security problems, passwords are still the incumbent authentication method for virtually all online services. To remedy the situation, end-users are very often referred to password managers as a solution to the password reuse and password weakness problems. However, to date the actual impact of password managers on password security and reuse has not been studied systematically. In this paper, we provide the first large-scale study of the password managers' influence on users' real-life passwords. From 476 participants of an online survey on users' password creation and management strategies, we recruit 170 participants that allowed us to monitor their passwords in-situ through a browser plugin. In contrast to prior work, we collect the passwords' entry methods (e.g., human or password manager) in addition to the passwords and their metrics. Based on our collected data and our survey, we gain a more complete picture of the factors that influence our participants' passwords' strength and reuse. We quantify for the first time that password managers indeed benefit the password strength and uniqueness, however, also our results also suggest that those benefits depend on the users' strategies and that managers without password generators rather aggravate the existing problems.

Patrick Speicher, Marcel Steinmetz, Jörg Hoffmann et al.

Penetration testing is a well-established practical concept for the identification of potentially exploitable security weaknesses and an important component of a security audit. Providing a holistic security assessment for networks consisting of several hundreds hosts is hardly feasible though without some sort of mechanization. Mitigation, prioritizing counter-measures subject to a given budget, currently lacks a solid theoretical understanding and is hence more art than science. In this work, we propose the first approach for conducting comprehensive what-if analyses in order to reason about mitigation in a conceptually well-founded manner. To evaluate and compare mitigation strategies, we use simulated penetration testing, i.e., automated attack-finding, based on a network model to which a subset of a given set of mitigation actions, e.g., changes to the network topology, system updates, configuration changes etc. is applied. Using Stackelberg planning, we determine optimal combinations that minimize the maximal attacker success (similar to a Stackelberg game), and thus provide a well-founded basis for a holistic mitigation strategy. We show that these Stackelberg planning models can largely be derived from network scan, public vulnerability databases and manual inspection with various degrees of automation and detail, and we simulate mitigation analysis on networks of different size and vulnerability.

Milivoj Simeonovski, Fabian Bendun, Muhammad Rizwan Asghar et al.

Search engines are the prevalently used tools to collect information about individuals on the Internet. Search results typically comprise a variety of sources that contain personal information -- either intentionally released by the person herself, or unintentionally leaked or published by third parties, often with detrimental effects on the individual's privacy. To grant individuals the ability to regain control over their disseminated personal information, the European Court of Justice recently ruled that EU citizens have a right to be forgotten in the sense that indexing systems, must offer them technical means to request removal of links from search results that point to sources violating their data protection rights. As of now, these technical means consist of a web form that requires a user to manually identify all relevant links upfront and to insert them into the web form, followed by a manual evaluation by employees of the indexing system to assess if the request is eligible and lawful. We propose a universal framework Oblivion to support the automation of the right to be forgotten in a scalable, provable and privacy-preserving manner. First, Oblivion enables a user to automatically find and tag her disseminated personal information using natural language processing and image recognition techniques and file a request in a privacy-preserving manner. Second, Oblivion provides indexing systems with an automated and provable eligibility mechanism, asserting that the author of a request is indeed affected by an online resource. The automated ligibility proof ensures censorship-resistance so that only legitimately affected individuals can request the removal of corresponding links from search results. We have conducted comprehensive evaluations, showing that Oblivion is capable of handling 278 removal requests per second, and is hence suitable for large-scale deployment.