Research Papers

An Adversarial Perspective on Machine Unlearning for AI Safety

Jakub Łucki, Boyi Wei, Yangsibo Huang +3

Large language models are finetuned to refuse questions about hazardous knowledge, but these protections can often be bypassed. Unlearning methods aim at completely removing hazardous capabilities from models and make them inaccessible to adversaries. This work challenges the fundamental differences between unlearning and traditional safety post-training from an adversarial perspective. We demonstrate that existing jailbreak methods, previously reported as ineffective against unlearning, can be successful when applied carefully. Furthermore, we develop a variety of adaptive methods that recover most supposedly unlearned capabilities. For instance, we show that finetuning on 10 unrelated examples or removing specific directions in the activation space can recover most hazardous capabilities for models edited with RMU, a state-of-the-art unlearning method. Our findings challenge the robustness of current unlearning approaches and question their advantages over safety training.

Adversarial examples mislead deep neural networks with imperceptible perturbations and have brought significant threats to deep learning. An important aspect is their transferability, which refers to their ability to deceive other models, thus enabling attacks in the black-box setting. Though various methods have been proposed to boost transferability, the performance still falls short compared with white-box attacks. In this work, we observe that existing input transformation based attacks, one of the mainstream transfer-based attacks, result in different attention heatmaps on various models, which might limit the transferability. We also find that breaking the intrinsic relation of the image can disrupt the attention heatmap of the original image. Based on this finding, we propose a novel input transformation based attack called block shuffle and rotation (BSR). Specifically, BSR splits the input image into several blocks, then randomly shuffles and rotates these blocks to construct a set of new images for gradient calculation. Empirical evaluations on the ImageNet dataset demonstrate that BSR could achieve significantly better transferability than the existing input transformation based methods under single-model and ensemble-model settings. Combining BSR with the current input transformation method can further improve the transferability, which significantly outperforms the state-of-the-art methods. Code is available at https://github.com/Trustworthy-AI-Group/BSR

Unbiased Watermark for Large Language Models

Zhengmian Hu, Lichang Chen, Xidong Wu +3

The recent advancements in large language models (LLMs) have sparked a growing apprehension regarding the potential misuse. One approach to mitigating this risk is to incorporate watermarking techniques into LLMs, allowing for the tracking and attribution of model outputs. This study examines a crucial aspect of watermarking: how significantly watermarks impact the quality of model-generated outputs. Previous studies have suggested a trade-off between watermark strength and output quality. However, our research demonstrates that it is possible to integrate watermarks without affecting the output probability distribution with appropriate implementation. We refer to this type of watermark as an unbiased watermark. This has significant implications for the use of LLMs, as it becomes impossible for users to discern whether a service provider has incorporated watermarks or not. Furthermore, the presence of watermarks does not compromise the performance of the model in downstream tasks, ensuring that the overall utility of the language model is preserved. Our findings contribute to the ongoing discussion around responsible AI development, suggesting that unbiased watermarks can serve as an effective means of tracking and attributing model outputs without sacrificing output quality.

Since the release of the original CheXpert paper five years ago, CheXpert has become one of the most widely used and cited clinical AI datasets. The emergence of vision language models has sparked an increase in demands for sharing reports linked to CheXpert images, along with a growing interest among AI fairness researchers in obtaining demographic data. To address this, CheXpert Plus serves as a new collection of radiology data sources, made publicly available to enhance the scaling, performance, robustness, and fairness of models for all subsequent machine learning tasks in the field of radiology. CheXpert Plus is the largest text dataset publicly released in radiology, with a total of 36 million text tokens, including 13 million impression tokens. To the best of our knowledge, it represents the largest text de-identification effort in radiology, with almost 1 million PHI spans anonymized. It is only the second time that a large-scale English paired dataset has been released in radiology, thereby enabling, for the first time, cross-institution training at scale. All reports are paired with high-quality images in DICOM format, along with numerous image and patient metadata covering various clinical and socio-economic groups, as well as many pathology labels and RadGraph annotations. We hope this dataset will boost research for AI models that can further assist radiologists and help improve medical care. Data is available at the following URL: https://stanfordaimi.azurewebsites.net/datasets/5158c524-d3ab-4e02-96e9-6ee9efc110a1 Models are available at the following URL: https://github.com/Stanford-AIMI/chexpert-plus

The advancement of secure communication and identity verification fields has significantly increased through the use of deep learning techniques for data hiding. By embedding information into a noise-tolerant signal such as audio, video, or images, digital watermarking and steganography techniques can be used to protect sensitive intellectual property and enable confidential communication, ensuring that the information embedded is only accessible to authorized parties. This survey provides an overview of recent developments in deep learning techniques deployed for data hiding, categorized systematically according to model architectures and noise injection methods. The objective functions, evaluation metrics, and datasets used for training these data hiding models are comprehensively summarised. Additionally, potential future research directions that unite digital watermarking and steganography on software engineering to enhance security and mitigate risks are suggested and deliberated. This contribution furthers the creation of a more trustworthy digital world and advances Responsible AI.

Ethics and Responsible AI Deployment

Petar Radanliev, Omar Santos

As Artificial Intelligence (AI) becomes more prevalent, protecting personal privacy is a critical ethical issue that must be addressed. This article explores the need for ethical AI systems that safeguard individual privacy while complying with ethical standards. By taking a multidisciplinary approach, the research examines innovative algorithmic techniques such as differential privacy, homomorphic encryption, federated learning, international regulatory frameworks, and ethical guidelines. The study concludes that these algorithms effectively enhance privacy protection while balancing the utility of AI with the need to protect personal data. The article emphasises the importance of a comprehensive approach that combines technological innovation with ethical and regulatory strategies to harness the power of AI in a way that respects and protects individual privacy.

The successful deployment of artificial intelligence (AI) in many domains from healthcare to hiring requires their responsible use, particularly in model explanations and privacy. Explainable artificial intelligence (XAI) provides more information to help users to understand model decisions, yet this additional knowledge exposes additional risks for privacy attacks. Hence, providing explanation harms privacy. We study this risk for image-based model inversion attacks and identified several attack architectures with increasing performance to reconstruct private image data from model explanations. We have developed several multi-modal transposed CNN architectures that achieve significantly higher inversion performance than using the target model prediction only. These XAI-aware inversion models were designed to exploit the spatial knowledge in image explanations. To understand which explanations have higher privacy risk, we analyzed how various explanation types and factors influence inversion performance. In spite of some models not providing explanations, we further demonstrate increased inversion performance even for non-explainable target models by exploiting explanations of surrogate models through attention transfer. This method first inverts an explanation from the target prediction, then reconstructs the target image. These threats highlight the urgent and significant privacy risks of explanations and calls attention for new privacy preservation techniques that balance the dual-requirement for AI explainability and privacy.

Fairness for Unobserved Characteristics: Insights from Technological Impacts on Queer Communities

Nenad Tomasev, Kevin R. McKee, Jackie Kay +1

Advances in algorithmic fairness have largely omitted sexual orientation and gender identity. We explore queer concerns in privacy, censorship, language, online safety, health, and employment to study the positive and negative effects of artificial intelligence on queer communities. These issues underscore the need for new directions in fairness research that take into account a multiplicity of considerations, from privacy preservation, context sensitivity and process fairness, to an awareness of sociotechnical impact and the increasingly important role of inclusive and participatory research processes. Most current approaches for algorithmic fairness assume that the target characteristics for fairness--frequently, race and legal gender--can be observed or recorded. Sexual orientation and gender identity are prototypical instances of unobserved characteristics, which are frequently missing, unknown or fundamentally unmeasurable. This paper highlights the importance of developing new approaches for algorithmic fairness that break away from the prevailing assumption of observed characteristics.

Ethics Sheet for Automatic Emotion Recognition and Sentiment Analysis

Saif M. Mohammad

The importance and pervasiveness of emotions in our lives makes affective computing a tremendously important and vibrant line of work. Systems for automatic emotion recognition (AER) and sentiment analysis can be facilitators of enormous progress (e.g., in improving public health and commerce) but also enablers of great harm (e.g., for suppressing dissidents and manipulating voters). Thus, it is imperative that the affective computing community actively engage with the ethical ramifications of their creations. In this paper, I have synthesized and organized information from AI Ethics and Emotion Recognition literature to present fifty ethical considerations relevant to AER. Notably, the sheet fleshes out assumptions hidden in how AER is commonly framed, and in the choices often made regarding the data, method, and evaluation. Special attention is paid to the implications of AER on privacy and social groups. Along the way, key recommendations are made for responsible AER. The objective of the sheet is to facilitate and encourage more thoughtfulness on why to automate, how to automate, and how to judge success well before the building of AER systems. Additionally, the sheet acts as a useful introductory document on emotion recognition (complementing survey articles).

The safety-critical scenarios of artificial intelligence (AI), such as autonomous driving, Internet of Things, smart healthcare, etc., have raised critical requirements of trustworthy AI to guarantee the privacy and security with reliable decisions. As a nascent branch for trustworthy AI, federated learning (FL) has been regarded as a promising privacy preserving framework for training a global AI model over collaborative devices. However, security challenges still exist in the FL framework, e.g., Byzantine attacks from malicious devices, and model tampering attacks from malicious server, which will degrade or destroy the accuracy of trained global AI model. In this paper, we shall propose a decentralized blockchain based FL (B-FL) architecture by using a secure global aggregation algorithm to resist malicious devices, and deploying practical Byzantine fault tolerance consensus protocol with high effectiveness and low energy consumption among multiple edge servers to prevent model tampering from the malicious server. However, to implement B-FL system at the network edge, multiple rounds of cross-validation in blockchain consensus protocol will induce long training latency. We thus formulate a network optimization problem that jointly considers bandwidth and power allocation for the minimization of long-term average training latency consisting of progressive learning rounds. We further propose to transform the network optimization problem as a Markov decision process and leverage the deep reinforcement learning based algorithm to provide high system performance with low computational complexity. Simulation results demonstrate that B-FL can resist malicious attacks from edge devices and servers, and the training latency of B-FL can be significantly reduced by deep reinforcement learning based algorithm compared with baseline algorithms.

EXplainable AI (XAI) methods have been proposed to interpret how a deep neural network predicts inputs through model saliency explanations that highlight the parts of the inputs deemed important to arrive a decision at a specific target. However, it remains challenging to quantify correctness of their interpretability as current evaluation approaches either require subjective input from humans or incur high computation cost with automated evaluation. In this paper, we propose backdoor trigger patterns--hidden malicious functionalities that cause misclassification--to automate the evaluation of saliency explanations. Our key observation is that triggers provide ground truth for inputs to evaluate whether the regions identified by an XAI method are truly relevant to its output. Since backdoor triggers are the most important features that cause deliberate misclassification, a robust XAI method should reveal their presence at inference time. We introduce three complementary metrics for systematic evaluation of explanations that an XAI method generates and evaluate seven state-of-the-art model-free and model-specific posthoc methods through 36 models trojaned with specifically crafted triggers using color, shape, texture, location, and size. We discovered six methods that use local explanation and feature relevance fail to completely highlight trigger regions, and only a model-free approach can uncover the entire trigger region.

Clio: Privacy-Preserving Insights into Real-World AI Use

Alex Tamkin, Miles McCain, Kunal Handa +18

How are AI assistants being used in the real world? While model providers in theory have a window into this impact via their users' data, both privacy concerns and practical challenges have made analyzing this data difficult. To address these issues, we present Clio (Claude insights and observations), a privacy-preserving platform that uses AI assistants themselves to analyze and surface aggregated usage patterns across millions of conversations, without the need for human reviewers to read raw conversations. We validate this can be done with a high degree of accuracy and privacy by conducting extensive evaluations. We demonstrate Clio's usefulness in two broad ways. First, we share insights about how models are being used in the real world from one million Claude.ai Free and Pro conversations, ranging from providing advice on hairstyles to providing guidance on Git operations and concepts. We also identify the most common high-level use cases on Claude.ai (coding, writing, and research tasks) as well as patterns that differ across languages (e.g., conversations in Japanese discuss elder care and aging populations at higher-than-typical rates). Second, we use Clio to make our systems safer by identifying coordinated attempts to abuse our systems, monitoring for unknown unknowns during critical periods like launches of new capabilities or major world events, and improving our existing monitoring systems. We also discuss the limitations of our approach, as well as risks and ethical concerns. By enabling analysis of real-world AI usage, Clio provides a scalable platform for empirically grounded AI safety and governance.