Safety backdoor attacks in large language models (LLMs) enable harmful behaviors to be stealthily triggered while evading detection during normal interactions. The high dimensionality of the trigger search space and the diverse range of potential malicious behaviors in LLMs make this a critical open problem. This paper presents BEEAR, a novel mitigation method based on a key insight: backdoor triggers induce a uniform drift in the model’s embedding space, irrespective of the trigger’s form or targeted behavior. Leveraging this observation, we introduce a bi-level optimization approach. The inner level identifies universal perturbations to the decoder’s embeddings that steer the model towards defender-defined unwanted behaviors; the outer level fine-tunes the model to reinforce safe behaviors against these perturbations. Our experiments demonstrate the effectiveness of this approach, reducing the success rate of safety backdoor attacks from over 95% to <1% for general harmful behaviors and from 47% to 0% for Sleeper Agents, without compromising the model’s helpfulness. Notably, our method relies only on defender-defined sets of safe and unwanted behaviors without any assumptions about the trigger location or attack mechanism. This work represents the first practical framework to counter safety backdoors in LLMs and provides a foundation for future advancements in AI safety and security.

Text summarization, a key natural language generation (NLG) task, is vital in various domains. However, the high cost of inaccurate summaries in risk-critical applications, particularly those involving human-in-the-loop decision-making, raises concerns about the reliability of uncertainty estimation on text summarization (UE-TS) evaluation methods. This concern stems from the dependency of uncertainty model metrics on diverse and potentially conflicting NLG metrics. To address this issue, we introduce a comprehensive UE-TS benchmark incorporating 31 NLG metrics across four dimensions. The benchmark evaluates the uncertainty estimation capabilities of two large language models and one pre-trained language model on three datasets, with human-annotation analysis incorporated where applicable. We also assess the performance of 14 common uncertainty estimation methods within this benchmark. Our findings emphasize the importance of considering multiple uncorrelated NLG metrics and diverse uncertainty estimation methods to ensure reliable and efficient evaluation of UE-TS techniques. Our code and data are available: https://github.com/he159ok/Benchmark-of-Uncertainty-Estimation-Methods-in-Text-Summarization.

Fact tracing seeks to identify specific training examples that serve as the knowledge source for a given query. Existing approaches to fact tracing rely on assessing the similarity between each training sample and the query along a certain dimension, such as lexical similarity, gradient, or embedding space. However, these methods fall short of effectively distinguishing between samples that are merely relevant and those that actually provide supportive evidence for the information sought by the query. This limitation often results in suboptimal effectiveness. Moreover, these approaches necessitate the examination of the similarity of individual training points for each query, imposing significant computational demands and creating a substantial barrier for practical applications. This paper introduces FASTTRACK, a novel approach that harnesses the capabilities of Large Language Models (LLMs) to validate supportive evidence for queries and at the same time clusters the training database towards a reduced extent for LLMs to trace facts. Our experiments show that FASTTRACK substantially outperforms existing methods in both accuracy and efficiency, achieving more than 100% improvement in F1 score over the state-of-the-art methods while being x33 faster than TracIn.

Large Language Models (LLMs) have shown remarkable capabilities in tasks such as summarization, arithmetic reasoning, and question answering. However, they encounter significant challenges in the domain of moral reasoning and ethical decision-making, especially in complex scenarios with multiple stakeholders. This paper introduces the Skin-in-the-Game (SKIG) framework, aimed at enhancing moral reasoning in LLMs by exploring decisions’ consequences from multiple stakeholder perspectives. The core components of the framework consist of simulating accountability for decisions, conducting empathy exercises on different stakeholders, and evaluating the risks associated with the impacts of potential actions. We study SKIG’s performance across various moral reasoning benchmarks with proprietary and open-source LLMs, and investigate its crucial components through extensive ablation analyses. Our framework exhibits marked improvements in performance compared to baselines across different language models and benchmarks.

Most traditional AI safety research views models as machines and centers on algorithm-focused attacks developed by security experts. As large language models (LLMs) become increasingly common and competent, non-expert users can also impose risks during daily interactions. Observing this, we shift the perspective, by treating LLMs as human-like communicators to examine the interplay between everyday language interaction and AI safety. Specifically, we study how to persuade LLMs to jailbreak them. First, we propose a persuasion taxonomy derived from decades of social science research. Then, we apply the taxonomy to automatically generate persuasive adversarial prompts (PAP) to jailbreak LLMs. Results show that persuasion significantly increases the jailbreak risk across all risk categories: PAP consistently achieves an attack success rate of over 92% on Llama-2-7b-Chat, GPT-3.5, and GPT-4 in 10 trials, surpassing recent algorithm-focused attacks. On the defense side, we explore various mechanisms against PAP, find a significant gap in existing defenses, and advocate for more fundamental solutions for AI safety.

With the increasing applications of language models, it has become crucial to protect these models from leaking private information. Previous work has attempted to tackle this challenge by training RNN-based language models with differential privacy guarantees. However, applying classical differential privacy to language models leads to poor model performance as the underlying privacy notion is over-pessimistic and provides undifferentiated protection for all tokens in the data. Given that the private information in natural language is sparse (for example, the bulk of an email might not carry personally identifiable information), we propose a new privacy notion, selective differential privacy, to provide rigorous privacy guarantees on the sensitive portion of the data to improve model utility. To realize such a new notion, we develop a corresponding privacy mechanism, Selective-DPSGD, for RNN-based language models. Besides language modeling, we also apply the method to a more concrete application – dialog systems. Experiments on both language modeling and dialog system building show that the proposed privacy-preserving mechanism achieves better utilities while remaining safe under various privacy attacks compared to the baselines. The data and code are released at https://github.com/wyshi/lm_privacy to facilitate future research.

Protecting large language models from privacy leakage is becoming increasingly crucial with their wide adoption in real-world products. Yet applying *differential privacy* (DP), a canonical notion with provable privacy guarantees for machine learning models, to those models remains challenging due to the trade-off between model utility and privacy loss. Utilizing the fact that sensitive information in language data tends to be sparse, Shi et al. (2021) formalized a DP notion extension called *Selective Differential Privacy* (SDP) to protect only the sensitive tokens defined by a policy function. However, their algorithm only works for RNN-based models. In this paper, we develop a novel framework, *Just Fine-tune Twice* (JFT), that achieves SDP for state-of-the-art large transformer-based models. Our method is easy to implement: it first fine-tunes the model with *redacted* in-domain data, and then fine-tunes it again with the *original* in-domain data using a private training mechanism. Furthermore, we study the scenario of imperfect implementation of policy functions that misses sensitive tokens and develop systematic methods to handle it. Experiments show that our method achieves strong utility compared to previous baselines. We also analyze the SDP privacy guarantee empirically with the canary insertion attack.