Activation Approximations Can Incur Safety Vulnerabilities Even in Aligned LLMs: Comprehensive Analysis and Defense

Large Language Models (LLMs) have showcased remarkable capabilities across various domains. Accompanying the evolving capabilities and expanding deployment scenarios of LLMs, their deployment challenges escalate due to their sheer scale and the advanced yet complex activation designs prevalent in notable model series, such as Llama, Gemma, and Mistral. These challenges have become particularly pronounced in resource-constrained deployment scenarios, where mitigating inference efficiency bottlenecks is imperative. Among various recent efforts, activation approximation has emerged as a promising avenue for pursuing inference efficiency, sometimes considered indispensable in applications such as private inference. Despite achieving substantial speedups with minimal impact on utility, even appearing sound and practical for real-world deployment, the safety implications of activation approximations remain unclear. In this work, we fill this critical gap in LLM safety by conducting the first systematic safety evaluation of activation approximations. Our safety vetting spans seven sota techniques across three popular categories, revealing consistent safety degradation across ten safety-aligned LLMs.

SecPE: Secure Prompt Ensembling for Private and Robust Large Language Models

With the growing popularity of LLMs among the general public users, privacy-preserving and adversarial robustness have become two pressing demands for LLM-based services, which have largely been pursued separately but rarely jointly. In this paper, to the best of our knowledge, we are among the first attempts towards robust and private LLM inference by tightly integrating two disconnected fields: private inference and prompt ensembling. The former protects users' privacy by encrypting inference data transmitted and processed by LLMs, while the latter enhances adversarial robustness by yielding an aggregated output from multiple prompted LLM responses. Although widely recognized as effective individually, private inference for prompt ensembling together entails new challenges that render the naive combination of existing techniques inefficient. To overcome the hurdles, we propose SecPE, which designs efficient fully homomorphic encryption (FHE) counterparts for the core algorithmic building blocks of prompt ensembling. We conduct extensive experiments on 8 tasks to evaluate the accuracy, robustness, and efficiency of SecPE. The results show that SecPE maintains high clean accuracy and offers better robustness at the expense of merely $2.5\%$ efficiency overhead compared to baseline private inference methods, indicating a satisfactory ``accuracy-robustness-efficiency'' tradeoff. For the efficiency of the encrypted Argmax operation that incurs major slowdown for prompt ensembling, SecPE is 35.4x faster than the state-of-the-art peers, which can be of independent interest beyond this work.

MemHunter: Automated and Verifiable Memorization Detection at Dataset-scale in LLMs

Large language models (LLMs) have been shown to memorize and reproduce content from their training data, raising significant privacy concerns, especially with web-scale datasets. Existing methods for detecting memorization are largely sample-specific, relying on manually crafted or discretely optimized memory-inducing prompts generated on a per-sample basis, which become impractical for dataset-level detection due to the prohibitive computational cost of iterating over all samples. In real-world scenarios, data owners may need to verify whether a susceptible LLM has memorized their dataset, particularly if the LLM may have collected the data from the web without authorization. To address this, we introduce \textit{MemHunter}, which trains a memory-inducing LLM and employs hypothesis testing to efficiently detect memorization at the dataset level, without requiring sample-specific memory inducing. Experiments on models such as Pythia and Llama-2 demonstrate that \textit{MemHunter} can extract up to 40\% more training data than existing methods under constrained time resources and reduce search time by up to 80\% when integrated as a plug-in. Crucially, \textit{MemHunter} is the first method capable of dataset-level memorization detection, providing an indispensable tool for assessing privacy risks in LLMs that are powered by vast web-sourced datasets.

Machine Unlearning in Forgettability Sequence

Machine unlearning (MU) is becoming a promising paradigm to achieve the "right to be forgotten", where the training trace of any chosen data points could be eliminated, while maintaining the model utility on general testing samples after unlearning. With the advancement of forgetting research, many fundamental open questions remain unanswered: do different samples exhibit varying levels of difficulty in being forgotten? Further, does the sequence in which samples are forgotten, determined by their respective difficulty levels, influence the performance of forgetting algorithms? In this paper, we identify key factor affecting unlearning difficulty and the performance of unlearning algorithms. We find that samples with higher privacy risks are more likely to be unlearning, indicating that the unlearning difficulty varies among different samples which motives a more precise unlearning mode. Built upon this insight, we propose a general unlearning framework, dubbed RSU, which consists of Ranking module and SeqUnlearn module.

Physical Backdoor: Towards Temperature-based Backdoor Attacks in the Physical World

Backdoor attacks have been well-studied in visible light object detection (VLOD) in recent years. However, VLOD can not effectively work in dark and temperature-sensitive scenarios. Instead, thermal infrared object detection (TIOD) is the most accessible and practical in such environments. In this paper, our team is the first to investigate the security vulnerabilities associated with TIOD in the context of backdoor attacks, spanning both the digital and physical realms. We introduce two novel types of backdoor attacks on TIOD, each offering unique capabilities: Object-affecting Attack and Range-affecting Attack. We conduct a comprehensive analysis of key factors influencing trigger design, which include temperature, size, material, and concealment. These factors, especially temperature, significantly impact the efficacy of backdoor attacks on TIOD. A thorough understanding of these factors will serve as a foundation for designing physical triggers and temperature controlling experiments. Our study includes extensive experiments conducted in both digital and physical environments. In the digital realm, we evaluate our approach using benchmark datasets for TIOD, achieving an Attack Success Rate (ASR) of up to 98.21%. In the physical realm, we test our approach in two real-world settings: a traffic intersection and a parking lot, using a thermal infrared camera. Here, we attain an ASR of up to 98.38%.

Clients Collaborate: Flexible Differentially Private Federated Learning with Guaranteed Improvement of Utility-Privacy Trade-off

To defend against privacy leakage of user data, differential privacy is widely used in federated learning, but it is not free. The addition of noise randomly disrupts the semantic integrity of the model and this disturbance accumulates with increased communication rounds. In this paper, we introduce a novel federated learning framework with rigorous privacy guarantees, named FedCEO, designed to strike a trade-off between model utility and user privacy by letting clients ''Collaborate with Each Other''. Specifically, we perform efficient tensor low-rank proximal optimization on stacked local model parameters at the server, demonstrating its capability to flexibly truncate high-frequency components in spectral space. This implies that our FedCEO can effectively recover the disrupted semantic information by smoothing the global semantic space for different privacy settings and continuous training processes. Moreover, we improve the SOTA utility-privacy trade-off bound by an order of $\sqrt{d}$, where $d$ is the input dimension. We illustrate our theoretical results with experiments on representative image datasets. It observes significant performance improvements and strict privacy guarantees under different privacy settings.

Cross-silo Federated Learning with Record-level Personalized Differential Privacy

Federated learning enhanced by differential privacy has emerged as a popular approach to better safeguard the privacy of client-side data by protecting clients' contributions during the training process. Existing solutions typically assume a uniform privacy budget for all records and provide one-size-fits-all solutions that may not be adequate to meet each record's privacy requirement. In this paper, we explore the uncharted territory of cross-silo FL with record-level personalized differential privacy. We devise a novel framework named rPDP-FL, employing a two-stage hybrid sampling scheme with both client-level sampling and non-uniform record-level sampling to accommodate varying privacy requirements. A critical and non-trivial problem is to select the ideal per-record sampling probability q given the personalized privacy budget {\epsilon}. We introduce a versatile solution named Simulation-CurveFitting, allowing us to uncover a significant insight into the nonlinear correlation between q and {\epsilon} and derive an elegant mathematical model to tackle the problem. Our evaluation demonstrates that our solution can provide significant performance gains over the baselines that do not consider personalized privacy preservation.

Contrastive Unlearning: A Contrastive Approach to Machine Unlearning

Machine unlearning aims to eliminate the influence of a subset of training samples (i.e., unlearning samples) from a trained model. Effectively and efficiently removing the unlearning samples without negatively impacting the overall model performance is still challenging. In this paper, we propose a contrastive unlearning framework, leveraging the concept of representation learning for more effective unlearning. It removes the influence of unlearning samples by contrasting their embeddings against the remaining samples so that they are pushed away from their original classes and pulled toward other classes. By directly optimizing the representation space, it effectively removes the influence of unlearning samples while maintaining the representations learned from the remaining samples. Experiments on a variety of datasets and models on both class unlearning and sample unlearning showed that contrastive unlearning achieves the best unlearning effects and efficiency with the lowest performance loss compared with the state-of-the-art algorithms.

Prompt Valuation Based on Shapley Values

Large language models (LLMs) excel on new tasks without additional training, simply by providing natural language prompts that demonstrate how the task should be performed. Prompt ensemble methods comprehensively harness the knowledge of LLMs while mitigating individual biases and errors and further enhancing performance. However, more prompts do not necessarily lead to better results, and not all prompts are beneficial. A small number of high-quality prompts often outperform many low-quality prompts. Currently, there is a lack of a suitable method for evaluating the impact of prompts on the results. In this paper, we utilize the Shapley value to fairly quantify the contributions of prompts, helping to identify beneficial or detrimental prompts, and potentially guiding prompt valuation in data markets. Through extensive experiments employing various ensemble methods and utility functions on diverse tasks, we validate the effectiveness of using the Shapley value method for prompts as it effectively distinguishes and quantifies the contributions of each prompt.

Signed Graph Neural Ordinary Differential Equation for Modeling Continuous-time Dynamics

Modeling continuous-time dynamics constitutes a foundational challenge, and uncovering inter-component correlations within complex systems holds promise for enhancing the efficacy of dynamic modeling. The prevailing approach of integrating graph neural networks with ordinary differential equations has demonstrated promising performance. However, they disregard the crucial signed information intrinsic to graphs, impeding their capacity to accurately capture real-world phenomena and leading to subpar outcomes. In response, we introduce a novel approach: a signed graph neural ordinary differential equation, adeptly addressing the limitations of miscapturing signed information. Our proposed solution boasts both flexibility and efficiency. To substantiate its effectiveness, we seamlessly integrate our devised strategies into three preeminent graph-based dynamic modeling frameworks: graph neural ordinary differential equations, graph neural controlled differential equations, and graph recurrent neural networks. Rigorous assessments encompass three intricate dynamic scenarios from physics and biology, as well as scrutiny across four authentic real-world traffic datasets. Remarkably outperforming the trio of baselines, empirical results underscore the substantial performance enhancements facilitated by our proposed approach.Our code can be found at https://github.com/beautyonce/SGODE.