Zhuo Zhang

Zhuo Zhang, Yuanhang Yang, Yong Dai et al. · tencent-ai

With increasing privacy concerns on data, recent studies have made significant progress using federated learning (FL) on privacy-sensitive natural language processing (NLP) tasks. Much literature suggests fully fine-tuning pre-trained language models (PLMs) in the FL paradigm can mitigate the data heterogeneity problem and close the performance gap with centralized training. However, large PLMs bring the curse of prohibitive communication overhead and local model adaptation costs for the FL system. To this end, we introduce various parameter-efficient tuning (PETuning) methods into federated learning. Specifically, we provide a holistic empirical study of representative PLMs tuning methods in FL. The experimental results cover the analysis of data heterogeneity levels, data scales, and different FL scenarios. Overall communication overhead can be significantly reduced by locally tuning and globally aggregating lightweight model parameters while maintaining acceptable performance in various FL settings. To facilitate the research of PETuning in FL, we also develop a federated tuning framework FedPETuning, which allows practitioners to exploit different PETuning methods under the FL training paradigm conveniently. The source code is available at \url{https://github.com/iezhuozhuo/FedETuning/tree/deltaTuning}.

Guanhong Tao, Zhenting Wang, Siyuan Cheng et al.

We conduct a systematic study of backdoor vulnerabilities in normally trained Deep Learning models. They are as dangerous as backdoors injected by data poisoning because both can be equally exploited. We leverage 20 different types of injected backdoor attacks in the literature as the guidance and study their correspondences in normally trained models, which we call natural backdoor vulnerabilities. We find that natural backdoors are widely existing, with most injected backdoor attacks having natural correspondences. We categorize these natural backdoors and propose a general detection framework. It finds 315 natural backdoors in the 56 normally trained models downloaded from the Internet, covering all the different categories, while existing scanners designed for injected backdoors can at most detect 65 backdoors. We also study the root causes and defense of natural backdoors.

Lu Yan, Zhuo Zhang, Guanhong Tao et al.

Backdoor attacks have emerged as a prominent threat to natural language processing (NLP) models, where the presence of specific triggers in the input can lead poisoned models to misclassify these inputs to predetermined target classes. Current detection mechanisms are limited by their inability to address more covert backdoor strategies, such as style-based attacks. In this work, we propose an innovative test-time poisoned sample detection framework that hinges on the interpretability of model predictions, grounded in the semantic meaning of inputs. We contend that triggers (e.g., infrequent words) are not supposed to fundamentally alter the underlying semantic meanings of poisoned samples as they want to stay stealthy. Based on this observation, we hypothesize that while the model's predictions for paraphrased clean samples should remain stable, predictions for poisoned samples should revert to their true labels upon the mutations applied to triggers during the paraphrasing process. We employ ChatGPT, a state-of-the-art large language model, as our paraphraser and formulate the trigger-removal task as a prompt engineering problem. We adopt fuzzing, a technique commonly used for unearthing software vulnerabilities, to discover optimal paraphrase prompts that can effectively eliminate triggers while concurrently maintaining input semantics. Experiments on 4 types of backdoor attacks, including the subtle style backdoors, and 4 distinct datasets demonstrate that our approach surpasses baseline methods, including STRIP, RAP, and ONION, in precision and recall.

Jiachen Yang, Zhuo Zhang, Yicheng Gong et al.

Data has now become a shortcoming of deep learning. Researchers in their own fields share the thinking that "deep neural networks might not always perform better when they eat more data," which still lacks experimental validation and a convincing guiding theory. Here to fill this lack, we design experiments from Identically Independent Distribution(IID) and Out of Distribution(OOD), which give powerful answers. For the purpose of guidance, based on the discussion of results, two theories are proposed: under IID condition, the amount of information determines the effectivity of each sample, the contribution of samples and difference between classes determine the amount of sample information and the amount of class information; under OOD condition, the cross-domain degree of samples determine the contributions, and the bias-fitting caused by irrelevant elements is a significant factor of cross-domain. The above theories provide guidance from the perspective of data, which can promote a wide range of practical applications of artificial intelligence.

Guanhong Tao, Yingqi Liu, Siyuan Cheng et al.

Pervasive backdoors are triggered by dynamic and pervasive input perturbations. They can be intentionally injected by attackers or naturally exist in normally trained models. They have a different nature from the traditional static and localized backdoors that can be triggered by perturbing a small input area with some fixed pattern, e.g., a patch with solid color. Existing defense techniques are highly effective for traditional backdoors. However, they may not work well for pervasive backdoors, especially regarding backdoor removal and model hardening. In this paper, we propose a novel model hardening technique against pervasive backdoors, including both natural and injected backdoors. We develop a general pervasive attack based on an encoder-decoder architecture enhanced with a special transformation layer. The attack can model a wide range of existing pervasive backdoor attacks and quantify them by class distances. As such, using the samples derived from our attack in adversarial training can harden a model against these backdoor vulnerabilities. Our evaluation on 9 datasets with 15 model structures shows that our technique can enlarge class distances by 59.65% on average with less than 1% accuracy degradation and no robustness loss, outperforming five hardening techniques such as adversarial training, universal adversarial training, MOTH, etc. It can reduce the attack success rate of six pervasive backdoor attacks from 99.06% to 1.94%, surpassing seven state-of-the-art backdoor removal techniques.

Ye Bai, Jie Li, Wenjing Han et al.

While transformers and their variant conformers show promising performance in speech recognition, the parameterized property leads to much memory cost during training and inference. Some works use cross-layer weight-sharing to reduce the parameters of the model. However, the inevitable loss of capacity harms the model performance. To address this issue, this paper proposes a parameter-efficient conformer via sharing sparsely-gated experts. Specifically, we use sparsely-gated mixture-of-experts (MoE) to extend the capacity of a conformer block without increasing computation. Then, the parameters of the grouped conformer blocks are shared so that the number of parameters is reduced. Next, to ensure the shared blocks with the flexibility of adapting representations at different levels, we design the MoE routers and normalization individually. Moreover, we use knowledge distillation to further improve the performance. Experimental results show that the proposed model achieves competitive performance with 1/3 of the parameters of the encoder, compared with the full-parameter model.

Zhuo Zhang, Guangyu Shen, Guanhong Tao et al.

Large Language Models (LLMs) are now widely used in various applications, making it crucial to align their ethical standards with human values. However, recent jail-breaking methods demonstrate that this alignment can be undermined using carefully constructed prompts. In our study, we reveal a new threat to LLM alignment when a bad actor has access to the model's output logits, a common feature in both open-source LLMs and many commercial LLM APIs (e.g., certain GPT models). It does not rely on crafting specific prompts. Instead, it exploits the fact that even when an LLM rejects a toxic request, a harmful response often hides deep in the output logits. By forcefully selecting lower-ranked output tokens during the auto-regressive generation process at a few critical output positions, we can compel the model to reveal these hidden responses. We term this process model interrogation. This approach differs from and outperforms jail-breaking methods, achieving 92% effectiveness compared to 62%, and is 10 to 20 times faster. The harmful content uncovered through our method is more relevant, complete, and clear. Additionally, it can complement jail-breaking strategies, with which results in further boosting attack performance. Our findings indicate that interrogation can extract toxic knowledge even from models specifically designed for coding tasks.

Guangyu Shen, Siyuan Cheng, Kaiyuan Zhang et al.

Large Language Models (LLMs) have become prevalent across diverse sectors, transforming human life with their extraordinary reasoning and comprehension abilities. As they find increased use in sensitive tasks, safety concerns have gained widespread attention. Extensive efforts have been dedicated to aligning LLMs with human moral principles to ensure their safe deployment. Despite their potential, recent research indicates aligned LLMs are prone to specialized jailbreaking prompts that bypass safety measures to elicit violent and harmful content. The intrinsic discrete nature and substantial scale of contemporary LLMs pose significant challenges in automatically generating diverse, efficient, and potent jailbreaking prompts, representing a continuous obstacle. In this paper, we introduce RIPPLE (Rapid Optimization via Subconscious Exploitation and Echopraxia), a novel optimization-based method inspired by two psychological concepts: subconsciousness and echopraxia, which describe the processes of the mind that occur without conscious awareness and the involuntary mimicry of actions, respectively. Evaluations across 6 open-source LLMs and 4 commercial LLM APIs show RIPPLE achieves an average Attack Success Rate of 91.5\%, outperforming five current methods by up to 47.0\% with an 8x reduction in overhead. Furthermore, it displays significant transferability and stealth, successfully evading established detection mechanisms. The code of our work is available at \url{https://github.com/SolidShen/RIPPLE_official/tree/official}

Zhuo Zhang, Jingyuan Zhang, Jintao Huang et al.

Instruction tuning has been identified as a crucial technique for optimizing the performance of large language models (LLMs) in generating human-aligned responses. Nonetheless, gathering diversified and superior-quality instruction data for such tuning presents notable obstacles, especially in domains with rigid privacy provisions. Federated instruction tuning (FedIT) has emerged as a promising solution, by consolidating collaborative training across multiple data owners, thereby resulting in a privacy-preserving learning model. However, FedIT encounters limitations such as scarcity of instructional data and risk of exposure to training data extraction attacks. In this paper, we propose a novel federated algorithm, FewFedPIT, designed to simultaneously enhance privacy protection and model performance of federated few-shot learning. FewFedPITcomprises three vital components on the client side: (1) synthetic data generation, which utilizes LLMs' in-context learning capacity to generate synthetic data autonomously, thus expanding the local database; (2) parameter isolation training, which individually updates the public parameters in the synthetic data and the private parameters in the local data, consequently mitigating the noise impact of the synthetic data; (3) local aggregation sharing, which mixes public and private parameters before uploading, effectively preventing data extraction attacks. Extensive experiments on three open-source datasets demonstrate the effectiveness of FewFedPITin, enhancing privacy preservation and improving federated few-shot performance.

Guangyu Shen, Siyuan Cheng, Xiangzhe Xu et al.

Large Language Models (LLMs) can acquire deceptive behaviors through backdoor attacks, where the model executes prohibited actions whenever secret triggers appear in the input. Existing safety training methods largely fail to address this vulnerability, due to the inherent difficulty of uncovering hidden triggers implanted in the model. Motivated by recent findings on LLMs' situational awareness, we propose a novel post-training framework that cultivates self-awareness of backdoor risks and enables models to articulate implanted triggers even when they are absent from the prompt. At its core, our approach introduces an inversion-inspired reinforcement learning framework that encourages models to introspectively reason about their own behaviors and reverse-engineer the triggers responsible for misaligned outputs. Guided by curated reward signals, this process transforms a poisoned model into one capable of precisely identifying its implanted trigger. Surprisingly, we observe that such backdoor self-awareness emerges abruptly within a short training window, resembling a phase transition in capability. Building on this emergent property, we further present two complementary defense strategies for mitigating and detecting backdoor threats. Experiments on five backdoor attacks, compared against six baseline methods, demonstrate that our approach has strong potential to improve the robustness of LLMs against backdoor risks. The code is available at LLM Backdoor Self-Awareness.

Zian Su, Xiangzhe Xu, Ziyang Huang et al.

Transformer based code models have impressive performance in many software engineering tasks. However, their effectiveness degrades when symbols are missing or not informative. The reason is that the model may not learn to pay attention to the right correlations/contexts without the help of symbols. We propose a new method to pre-train general code models when symbols are lacking. We observe that in such cases, programs degenerate to something written in a very primitive language. We hence propose to use program analysis to extract contexts a priori (instead of relying on symbols and masked language modeling as in vanilla models). We then leverage a novel attention masking method to only allow the model attending to these contexts, e.g., bi-directional program dependence transitive closures and token co-occurrences. In the meantime, the inherent self-attention mechanism is utilized to learn which of the allowed attentions are more important compared to others. To realize the idea, we enhance the vanilla tokenization and model architecture of a BERT model, construct and utilize attention masks, and introduce a new pre-training algorithm. We pre-train this BERT-like model from scratch, using a dataset of 26 million stripped binary functions with explicit program dependence information extracted by our tool. We apply the model in three downstream tasks: binary similarity, type inference, and malware family classification. Our pre-trained model can improve the SOTAs in these tasks from 53% to 64%, 49% to 60%, and 74% to 94%, respectively. It also substantially outperforms other general pre-training techniques of code understanding models.

Guangyu Shen, Yingqi Liu, Guanhong Tao et al.

We develop a novel optimization method for NLPbackdoor inversion. We leverage a dynamically reducing temperature coefficient in the softmax function to provide changing loss landscapes to the optimizer such that the process gradually focuses on the ground truth trigger, which is denoted as a one-hot value in a convex hull. Our method also features a temperature rollback mechanism to step away from local optimals, exploiting the observation that local optimals can be easily deter-mined in NLP trigger inversion (while not in general optimization). We evaluate the technique on over 1600 models (with roughly half of them having injected backdoors) on 3 prevailing NLP tasks, with 4 different backdoor attacks and 7 architectures. Our results show that the technique is able to effectively and efficiently detect and remove backdoors, outperforming 4 baseline methods.