While large language models (LLMs) have demonstrated increasing power, they have also called upon studies on their vulnerabilities. As representatives, jailbreak attacks can provoke harmful or unethical responses from LLMs, even after safety alignment. In this paper, we investigate a novel category of jailbreak attacks specifically designed to target the cognitive structure and processes of LLMs. Specifically, we analyze the safety vulnerability of LLMs in the face of 1) multilingual cognitive overload, 2) veiled expression, and 3) effect-to- cause reasoning. Different from previous jailbreak attacks, our proposed cognitive overload is a black-box attack with no need for knowledge of model architecture or access to model weights. Experiments conducted on AdvBench and MasterKey reveal that various LLMs, including both popular open-source model Llama 2 and the proprietary model ChatGPT, can be compromised through cognitive overload. Motivated by cognitive psychology work on managing cognitive load, we further investigate defending cognitive overload attack from two perspectives. Empirical studies show that our cognitive overload from three perspectives can jailbreak all studied LLMs successfully, while existing defense strategies can hardly mitigate the caused malicious uses effectively.

Language models are often at risk of diverse backdoor attacks, especially data poisoning. Thus, it is important to investigate defense solutions for addressing them. Existing backdoor defense methods mainly focus on backdoor attacks with explicit triggers, leaving a universal defense against various backdoor attacks with diverse triggers largely unexplored. In this paper, we propose an end-to-end ensemble-based backdoor defense framework, DPoE (Denoised Product-of-Experts), which is inspired by the shortcut nature of backdoor attacks, to defend various backdoor attacks. DPoE consists of two models: a shallow model that captures the backdoor shortcuts and a main model that is prevented from learning the shortcuts. To address the label flip caused by backdoor attackers, DPoE incorporates a denoising design. Experiments on three NLP tasks show that DPoE significantly improves the defense performance against various types of backdoor triggers including word-level, sentence-level, and syntactic triggers. Furthermore, DPoE is also effective under a more challenging but practical setting that mixes multiple types of triggers.

Textual backdoor attacks, characterized by subtle manipulations of input triggers and training dataset labels, pose significant threats to security-sensitive applications. The rise of advanced generative models, such as GPT-4, with their capacity for human-like rewriting, makes these attacks increasingly challenging to detect. In this study, we conduct an in-depth examination of black-box generative models as tools for backdoor attacks, thereby emphasizing the need for effective defense strategies. We propose BGMAttack, a novel framework that harnesses advanced generative models to execute stealthier backdoor attacks on text classifiers. Unlike prior approaches constrained by subpar generation quality, BGMAttack renders backdoor triggers more elusive to human cognition and advanced machine detection. A rigorous evaluation of attack effectiveness over four sentiment classification tasks, complemented by four human cognition stealthiness tests, reveals BGMAttack’s superior performance, achieving a state-of-the-art attack success rate of 97.35% on average while maintaining superior stealth compared to conventional methods. The dataset and code are available: https://github.com/JiazhaoLi/BGMAttack.

We investigate security concerns of the emergent instruction tuning paradigm, that models are trained on crowdsourced datasets with task instructions to achieve superior performance. Our studies demonstrate that an attacker can inject backdoors by issuing very few malicious instructions (~1000 tokens) and control model behavior through data poisoning, without even the need to modify data instances or labels themselves. Through such instruction attacks, the attacker can achieve over 90% attack success rate across four commonly used NLP datasets. As an empirical study on instruction attacks, we systematically evaluated unique perspectives of instruction attacks, such as poison transfer where poisoned models can transfer to 15 diverse generative datasets in a zero-shot manner; instruction transfer where attackers can directly apply poisoned instruction on many other datasets; and poison resistance to continual finetuning. Lastly, we show that RLHF and clean demonstrations might mitigate such backdoors to some degree. These findings highlight the need for more robust defenses against poisoning attacks in instruction-tuning models and underscore the importance of ensuring data quality in instruction crowdsourcing.

The exorbitant cost of training Large language models (LLMs) from scratch makes it essential to fingerprint the models to protect intellectual property via ownership authentication and to ensure downstream users and developers comply with their license terms (eg restricting commercial use). In this study, we present a pilot study on LLM fingerprinting as a form of very lightweight instruction tuning. Model publisher specifies a confidential private key and implants it as an instruction backdoor that causes the LLM to generate specific text when the key is present. Results on 11 popularly-used LLMs showed that this approach is lightweight and does not affect the normal behavior of the model. It also prevents publisher overclaim, maintains robustness against fingerprint guessing and parameter-efficient training, and supports multi-stage fingerprinting akin to MIT License.

This tutorial seeks to provide a systematic summary of risks and vulnerabilities in security, privacy and copyright aspects of large language models (LLMs), and most recent solutions to address those issues. We will discuss a broad thread of studies that try to answer the following questions: (i) How do we unravel the adversarial threats that attackers may leverage in the training time of LLMs, especially those that may exist in recent paradigms of instruction tuning and RLHF processes? (ii) How do we guard the LLMs against malicious attacks in inference time, such as attacks based on backdoors and jailbreaking? (iii) How do we ensure privacy protection of user information and LLM decisions for Language Model as-a-Service (LMaaS)? (iv) How do we protect the copyright of an LLM? (v) How do we detect and prevent cases where personal or confidential information is leaked during LLM training? (vi) How should we make policies to control against improper usage of LLM-generated content? In addition, will conclude the discussions by outlining emergent challenges in security, privacy and reliability of LLMs that deserve timely investigation by the community

Reinforcement Learning with Human Feedback (RLHF) is a methodology designed to align Large Language Models (LLMs) with human preferences, playing an important role in LLMs alignment. Despite its advantages, RLHF relies on human annotators to rank the text, which can introduce potential security vulnerabilities if any adversarial annotator (i.e., attackers) manipulates the ranking score by up-ranking any malicious text to steer the LLM adversarially. To assess the red-teaming of RLHF against human preference data poisoning, we propose RankPoison, a poisoning attack method on candidates’ selection of preference rank flipping to reach certain malicious behaviors (e.g., generating longer sequences, which can increase the computational cost). With poisoned dataset generated by RankPoison, we can perform poisoning attacks on LLMs to generate longer tokens without hurting the original safety alignment performance. Moreover, applying RankPoison, we also successfully implement a backdoor attack where LLMs can generate longer answers under questions with the trigger word. Our findings highlight critical security challenges in RLHF, underscoring the necessity for more robust alignment methods for LLMs.

Textual backdoor attack, as a novel attack model, has been shown to be effective in adding a backdoor to the model during training. Defending against such backdoor attacks has become urgent and important. In this paper, we propose AttDef, an efficient attribution-based pipeline to defend against two insertion-based poisoning attacks, BadNL and InSent. Specifically, we regard the tokens with larger attribution scores as potential triggers since larger attribution words contribute more to the false prediction results and therefore are more likely to be poison triggers. Additionally, we further utilize an external pre-trained language model to distinguish whether input is poisoned or not. We show that our proposed method can generalize sufficiently well in two common attack scenarios (poisoning training data and testing data), which consistently improves previous methods. For instance, AttDef can successfully mitigate both attacks with an average accuracy of 79.97% (56.59% up) and 48.34% (3.99% up) under pre-training and post-training attack defense respectively, achieving the new state-of-the-art performance on prediction recovery over four benchmark datasets.

In this paper, we propose a hierarchical contrastive learning framework, HiCL, which considers local segment-level and global sequence-level relationships to improve training efficiency and effectiveness. Traditional methods typically encode a sequence in its entirety for contrast with others, often neglecting local representation learning, leading to challenges in generalizing to shorter texts. Conversely, HiCL improves its effectiveness by dividing the sequence into several segments and employing both local and global contrastive learning to model segment-level and sequence-level relationships. Further, considering the quadratic time complexity of transformers over input tokens, HiCL boosts training efficiency by first encoding short segments and then aggregating them to obtain the sequence representation. Extensive experiments show that HiCL enhances the prior top-performing SNCSE model across seven extensively evaluated STS tasks, with an average increase of +0.2% observed on BERT_large and +0.44% on RoBERTa_large.

Augmenting pretrained language models (LMs) with a vision encoder (e.g., Flamingo) has obtained state-of-the-art results in image-to-text generation. However, these models store all the knowledge within their parameters, thus often requiring enormous model parameters to model the abundant visual concepts and very rich text descriptions. Additionally, they are inefficient in incorporating new data, requiring a computational-expensive fine-tuning process. In this work, we introduce a Retrieval-augmented Visual Language Model, Re-ViLM, built upon the Flamingo, that supports retrieving the relevant knowledge from the external database for zero and in-context few-shot image-to-text generations. By storing certain knowledge explicitly in the external database, our approach reduces the number of model parameters and can easily accommodate new data during evaluation by simply updating the database. We also construct an interleaved image and text data that facilitates in-context few-shot learning capabilities.We demonstrate that Re-ViLM significantly boosts performance for image-to-text generation tasks, especially for zero-shot and few-shot generation in out-of-domain settings with 4x less parameters compared with baseline methods.

Large decoder-only language models (LMs) can be largely improved in terms of perplexity by retrieval (e.g., RETRO), but its impact on text generation quality and downstream task accuracy is unclear. Thus, it is still an open question: shall we pretrain large autoregressive LMs with retrieval? To answer it, we perform a comprehensive study on a scalable pre-trained retrieval-augmented LM (i.e., RETRO) compared with standard GPT and retrieval-augmented GPT incorporated at fine-tuning or inference stages. We first provide the recipe to reproduce RETRO up to 9.5B parameters while retrieving a text corpus with 330B tokens. Based on that, we have the following novel findings: i) RETRO outperforms GPT on text generation with much less degeneration (i.e., repetition), moderately higher factual accuracy, and slightly lower toxicity with a nontoxic retrieval database. ii) On the LM Evaluation Harness benchmark, RETRO largely outperforms GPT on knowledge-intensive tasks, but is on par with GPT on other tasks. Furthermore, we introduce a simple variant of the model, RETRO++, which largely improves open-domain QA results of original RETRO (e.g., EM score +8.6 on Natural Question) and significantly outperforms retrieval-augmented GPT across different model sizes. Our findings highlight the promising direction of pretraining autoregressive LMs with retrieval as future foundation models. We release our implementation at: https://github.com/NVIDIA/Megatron-LM/tree/main/tools/retro.