Comprehensive Assessment of Jailbreak Attacks Against LLMs

Misuse of the Large Language Models (LLMs) has raised widespread concern. To address this issue, safeguards have been taken to ensure that LLMs align with social ethics. However, recent findings have revealed an unsettling vulnerability bypassing the safeguards of LLMs, known as jailbreak attacks. By applying techniques, such as employing role-playing scenarios, adversarial examples, or subtle subversion of safety objectives as a prompt, LLMs can produce an inappropriate or even harmful response. While researchers have studied several categories of jailbreak attacks, they have done so in isolation. To fill this gap, we present the first large-scale measurement of various jailbreak attack methods. We concentrate on 13 cutting-edge jailbreak methods from four categories, 160 questions from 16 violation categories, and six popular LLMs. Our extensive experimental results demonstrate that the optimized jailbreak prompts consistently achieve the highest attack success rates, as well as exhibit robustness across different LLMs. Some jailbreak prompt datasets, available from the Internet, can also achieve high attack success rates on many LLMs, such as ChatGLM3, GPT-3.5, and PaLM2. Despite the claims from many organizations regarding the coverage of violation categories in their policies, the attack success rates from these categories remain high, indicating the challenges of effectively aligning LLM policies and the ability to counter jailbreak attacks. We also discuss the trade-off between the attack performance and efficiency, as well as show that the transferability of the jailbreak prompts is still viable, becoming an option for black-box models. Overall, our research highlights the necessity of evaluating different jailbreak methods. We hope our study can provide insights for future research on jailbreak attacks and serve as a benchmark tool for evaluating them for practitioners.

SecurityNet: Assessing Machine Learning Vulnerabilities on Public Models

While advanced machine learning (ML) models are deployed in numerous real-world applications, previous works demonstrate these models have security and privacy vulnerabilities. Various empirical research has been done in this field. However, most of the experiments are performed on target ML models trained by the security researchers themselves. Due to the high computational resource requirement for training advanced models with complex architectures, researchers generally choose to train a few target models using relatively simple architectures on typical experiment datasets. We argue that to understand ML models' vulnerabilities comprehensively, experiments should be performed on a large set of models trained with various purposes (not just the purpose of evaluating ML attacks and defenses). To this end, we propose using publicly available models with weights from the Internet (public models) for evaluating attacks and defenses on ML models. We establish a database, namely SecurityNet, containing 910 annotated image classification models. We then analyze the effectiveness of several representative attacks/defenses, including model stealing attacks, membership inference attacks, and backdoor detection on these public models. Our evaluation empirically shows the performance of these attacks/defenses can vary significantly on public models compared to self-trained models. We share SecurityNet with the research community. and advocate researchers to perform experiments on public models to better demonstrate their proposed methods' effectiveness in the future.

IDOL: Indicator-oriented Logic Pre-training for Logical Reasoning

In the field of machine reading comprehension (MRC), existing systems have surpassed the average performance of human beings in many tasks like SQuAD. However, there is still a long way to go when it comes to logical reasoning. Although some methods for it have been put forward, they either are designed in a quite complicated way or rely too much on external structures. In this paper, we proposed IDOL (InDicator-Oriented Logic Pre-training), an easy-to-understand but highly effective further pre-training task which logically strengthens the pre-trained models with the help of 6 types of logical indicators and a logically rich dataset LGP (LoGic Pre-training). IDOL achieves state-of-the-art performance on ReClor and LogiQA, the two most representative benchmarks in logical reasoning MRC, and is proven to be capable of generalizing to different pre-trained models and other types of MRC benchmarks like RACE and SQuAD 2.0 while keeping competitive general language understanding ability through testing on tasks in GLUE. Besides, at the beginning of the era of large language models, we take several of them like ChatGPT into comparison and find that IDOL still shows its advantage.

Efficient and Effective Text Encoding for Chinese LLaMA and Alpaca

Large Language Models (LLMs), such as ChatGPT and GPT-4, have revolutionized natural language processing research and demonstrated potential in Artificial General Intelligence (AGI). However, the expensive training and deployment of LLMs present challenges to transparent and open academic research. To address these issues, this project open-sources the Chinese LLaMA and Alpaca large models, emphasizing instruction fine-tuning. We expand the original LLaMA's Chinese vocabulary by adding 20K Chinese tokens, increasing encoding efficiency and enhancing basic semantic understanding. By incorporating secondary pre-training using Chinese data and fine-tuning with Chinese instruction data, we substantially improve the models' comprehension and execution of instructions. Our pilot study serves as a foundation for researchers adapting LLaMA and Alpaca models to other languages. Resources are made publicly available through GitHub, fostering open research in the Chinese NLP community and beyond. GitHub repository: https://github.com/ymcui/Chinese-LLaMA-Alpaca

MiniRBT: A Two-stage Distilled Small Chinese Pre-trained Model

In natural language processing, pre-trained language models have become essential infrastructures. However, these models often suffer from issues such as large size, long inference time, and challenging deployment. Moreover, most mainstream pre-trained models focus on English, and there are insufficient studies on small Chinese pre-trained models. In this paper, we introduce MiniRBT, a small Chinese pre-trained model that aims to advance research in Chinese natural language processing. MiniRBT employs a narrow and deep student model and incorporates whole word masking and two-stage distillation during pre-training to make it well-suited for most downstream tasks. Our experiments on machine reading comprehension and text classification tasks reveal that MiniRBT achieves 94% performance relative to RoBERTa, while providing a 6.8x speedup, demonstrating its effectiveness and efficiency.

From Visual Prompt Learning to Zero-Shot Transfer: Mapping Is All You Need

Visual prompt learning, as a newly emerged technique, leverages the knowledge learned by a large-scale pre-trained model and adapts it to downstream tasks through the usage of prompts. While previous research has focused on designing effective prompts, in this work, we argue that compared to prompt design, a good mapping strategy matters more. In this sense, we propose SeMap, a more effective mapping using the semantic alignment between the pre-trained model's knowledge and the downstream task. Our experimental results show that SeMap can largely boost the performance of visual prompt learning. Moreover, our experiments show that SeMap is capable of achieving competitive zero-shot transfer, indicating that it can perform the downstream task without any fine-tuning on the corresponding dataset. This demonstrates the potential of our proposed method to be used in a broader range of applications where the zero-shot transfer is desired. Results suggest that our proposed SeMap could lead to significant advancements in both visual prompt learning and zero-shot transfer. We hope with SeMap, we can help the community move forward to more efficient and lightweight utilization of large vision models.

Gradient-based Intra-attention Pruning on Pre-trained Language Models

Pre-trained language models achieve superior performance, but they are computationally expensive due to their large size. Techniques such as pruning and knowledge distillation (KD) have been developed to reduce their size and latency. In most structural pruning methods, the pruning units, such as attention heads and feed-forward hidden dimensions, only span a small model structure space and limit the structures that the pruning algorithm can explore. In this work, we propose Gradient-based Intra-attention pruning (GRAIN), which inspects fine intra-attention structures, and allows different heads to have different sizes. Intra-attention pruning greatly expands the searching space of model structures and yields highly heterogeneous structures. We further propose structure regularization to encourage generating more regular structures, which achieves higher speedups than heterogeneous ones. We also integrate KD into the pruning process with a gradient separation strategy to reduce the interference of KD with the pruning process. GRAIN is evaluated on a variety of tasks. Results show that it notably outperforms other methods at the same or similar model size. Even under extreme compression where only $3\%$ weights in transformers remain, the pruned model is still competitive.

Data Poisoning Attacks Against Multimodal Encoders

Traditional machine learning (ML) models usually rely on large-scale labeled datasets to achieve strong performance. However, such labeled datasets are often challenging and expensive to obtain. Also, the predefined categories limit the model's ability to generalize to other visual concepts as additional labeled data is required. On the contrary, the newly emerged multimodal model, which contains both visual and linguistic modalities, learns the concept of images from the raw text. It is a promising way to solve the above problems as it can use easy-to-collect image-text pairs to construct the training dataset and the raw texts contain almost unlimited categories according to their semantics. However, learning from a large-scale unlabeled dataset also exposes the model to the risk of potential poisoning attacks, whereby the adversary aims to perturb the model's training dataset to trigger malicious behaviors in it. Previous work mainly focuses on the visual modality. In this paper, we instead focus on answering two questions: (1) Is the linguistic modality also vulnerable to poisoning attacks? and (2) Which modality is most vulnerable? To answer the two questions, we conduct three types of poisoning attacks against CLIP, the most representative multimodal contrastive learning framework. Extensive evaluations on different datasets and model architectures show that all three attacks can perform well on the linguistic modality with only a relatively low poisoning rate and limited epochs. Also, we observe that the poisoning effect differs between different modalities, i.e., with lower MinRank in the visual modality and with higher Hit@K when K is small in the linguistic modality. To mitigate the attacks, we propose both pre-training and post-training defenses. We empirically show that both defenses can significantly reduce the attack performance while preserving the model's utility.

HIT at SemEval-2022 Task 2: Pre-trained Language Model for Idioms Detection

The same multi-word expressions may have different meanings in different sentences. They can be mainly divided into two categories, which are literal meaning and idiomatic meaning. Non-contextual-based methods perform poorly on this problem, and we need contextual embedding to understand the idiomatic meaning of multi-word expressions correctly. We use a pre-trained language model, which can provide a context-aware sentence embedding, to detect whether multi-word expression in the sentence is idiomatic usage.

HFL at SemEval-2022 Task 8: A Linguistics-inspired Regression Model with Data Augmentation for Multilingual News Similarity

This paper describes our system designed for SemEval-2022 Task 8: Multilingual News Article Similarity. We proposed a linguistics-inspired model trained with a few task-specific strategies. The main techniques of our system are: 1) data augmentation, 2) multi-label loss, 3) adapted R-Drop, 4) samples reconstruction with the head-tail combination. We also present a brief analysis of some negative methods like two-tower architecture. Our system ranked 1st on the leaderboard while achieving a Pearson's Correlation Coefficient of 0.818 on the official evaluation set.