GradEscape: A Gradient-Based Evader Against AI-Generated Text Detectors

In this paper, we introduce GradEscape, the first gradient-based evader designed to attack AI-generated text (AIGT) detectors. GradEscape overcomes the undifferentiable computation problem, caused by the discrete nature of text, by introducing a novel approach to construct weighted embeddings for the detector input. It then updates the evader model parameters using feedback from victim detectors, achieving high attack success with minimal text modification. To address the issue of tokenizer mismatch between the evader and the detector, we introduce a warm-started evader method, enabling GradEscape to adapt to detectors across any language model architecture. Moreover, we employ novel tokenizer inference and model extraction techniques, facilitating effective evasion even in query-only access. We evaluate GradEscape on four datasets and three widely-used language models, benchmarking it against four state-of-the-art AIGT evaders. Experimental results demonstrate that GradEscape outperforms existing evaders in various scenarios, including with an 11B paraphrase model, while utilizing only 139M parameters. We have successfully applied GradEscape to two real-world commercial AIGT detectors. Our analysis reveals that the primary vulnerability stems from disparity in text expression styles within the training data. We also propose a potential defense strategy to mitigate the threat of AIGT evaders. We open-source our GradEscape for developing more robust AIGT detectors.

VModA: An Effective Framework for Adaptive NSFW Image Moderation

Not Safe/Suitable for Work (NSFW) content is rampant on social networks and poses serious harm to citizens, especially minors. Current detection methods mainly rely on deep learning-based image recognition and classification. However, NSFW images are now presented in increasingly sophisticated ways, often using image details and complex semantics to obscure their true nature or attract more views. Although still understandable to humans, these images often evade existing detection methods, posing a significant threat. Further complicating the issue, varying regulations across platforms and regions create additional challenges for effective moderation, leading to detection bias and reduced accuracy. To address this, we propose VModA, a general and effective framework that adapts to diverse moderation rules and handles complex, semantically rich NSFW content across categories. Experimental results show that VModA significantly outperforms existing methods, achieving up to a 54.3% accuracy improvement across NSFW types, including those with complex semantics. Further experiments demonstrate that our method exhibits strong adaptability across categories, scenarios, and base VLMs. We also identified inconsistent and controversial label samples in public NSFW benchmark datasets, re-annotated them, and submitted corrections to the original maintainers. Two datasets have confirmed the updates so far. Additionally, we evaluate VModA in real-world scenarios to demonstrate its practical effectiveness.

MentalMAC: Enhancing Large Language Models for Detecting Mental Manipulation via Multi-Task Anti-Curriculum Distillation

Mental manipulation is a subtle yet pervasive form of psychological abuse that poses serious threats to mental health. Its covert nature and the complexity of manipulation strategies make it challenging to detect, even for state-of-the-art large language models (LLMs). This concealment also hinders the manual collection of large-scale, high-quality annotations essential for training effective models. Although recent efforts have sought to improve LLM's performance on this task, progress remains limited due to the scarcity of real-world annotated datasets. To address these challenges, we propose MentalMAC, a multi-task anti-curriculum distillation method that enhances LLMs' ability to detect mental manipulation in multi-turn dialogue. Our approach includes: (i) EvoSA, an unsupervised data expansion method based on evolutionary operations and speech act theory; (ii) teacher-model-generated multi-task supervision; and (iii) progressive knowledge distillation from complex to simpler tasks. We then constructed the ReaMent dataset with 5,000 real-world dialogue samples, using a MentalMAC-distilled model to assist human annotation. Vast experiments demonstrate that our method significantly narrows the gap between student and teacher models and outperforms competitive LLMs across key evaluation metrics. All code, datasets, and checkpoints will be released upon paper acceptance. Warning: This paper contains content that may be offensive to readers.

DC-SGD: Differentially Private SGD with Dynamic Clipping through Gradient Norm Distribution Estimation

Differentially Private Stochastic Gradient Descent (DP-SGD) is a widely adopted technique for privacy-preserving deep learning. A critical challenge in DP-SGD is selecting the optimal clipping threshold C, which involves balancing the trade-off between clipping bias and noise magnitude, incurring substantial privacy and computing overhead during hyperparameter tuning. In this paper, we propose Dynamic Clipping DP-SGD (DC-SGD), a framework that leverages differentially private histograms to estimate gradient norm distributions and dynamically adjust the clipping threshold C. Our framework includes two novel mechanisms: DC-SGD-P and DC-SGD-E. DC-SGD-P adjusts the clipping threshold based on a percentile of gradient norms, while DC-SGD-E minimizes the expected squared error of gradients to optimize C. These dynamic adjustments significantly reduce the burden of hyperparameter tuning C. The extensive experiments on various deep learning tasks, including image classification and natural language processing, show that our proposed dynamic algorithms achieve up to 9 times acceleration on hyperparameter tuning than DP-SGD. And DC-SGD-E can achieve an accuracy improvement of 10.62% on CIFAR10 than DP-SGD under the same privacy budget of hyperparameter tuning. We conduct rigorous theoretical privacy and convergence analyses, showing that our methods seamlessly integrate with the Adam optimizer. Our results highlight the robust performance and efficiency of DC-SGD, offering a practical solution for differentially private deep learning with reduced computational overhead and enhanced privacy guarantees.

Dialogue Injection Attack: Jailbreaking LLMs through Context Manipulation

Large language models (LLMs) have demonstrated significant utility in a wide range of applications; however, their deployment is plagued by security vulnerabilities, notably jailbreak attacks. These attacks manipulate LLMs to generate harmful or unethical content by crafting adversarial prompts. While much of the current research on jailbreak attacks has focused on single-turn interactions, it has largely overlooked the impact of historical dialogues on model behavior. In this paper, we introduce a novel jailbreak paradigm, Dialogue Injection Attack (DIA), which leverages the dialogue history to enhance the success rates of such attacks. DIA operates in a black-box setting, requiring only access to the chat API or knowledge of the LLM's chat template. We propose two methods for constructing adversarial historical dialogues: one adapts gray-box prefilling attacks, and the other exploits deferred responses. Our experiments show that DIA achieves state-of-the-art attack success rates on recent LLMs, including Llama-3.1 and GPT-4o. Additionally, we demonstrate that DIA can bypass 5 different defense mechanisms, highlighting its robustness and effectiveness.

R.R.: Unveiling LLM Training Privacy through Recollection and Ranking

Large Language Models (LLMs) pose significant privacy risks, potentially leaking training data due to implicit memorization. Existing privacy attacks primarily focus on membership inference attacks (MIAs) or data extraction attacks, but reconstructing specific personally identifiable information (PII) in LLM's training data remains challenging. In this paper, we propose R.R. (Recollect and Rank), a novel two-step privacy stealing attack that enables attackers to reconstruct PII entities from scrubbed training data where the PII entities have been masked. In the first stage, we introduce a prompt paradigm named recollection, which instructs the LLM to repeat a masked text but fill in masks. Then we can use PII identifiers to extract recollected PII candidates. In the second stage, we design a new criterion to score each PII candidate and rank them. Motivated by membership inference, we leverage the reference model as a calibration to our criterion. Experiments across three popular PII datasets demonstrate that the R.R. achieves better PII identical performance compared to baselines. These results highlight the vulnerability of LLMs to PII leakage even when training data has been scrubbed. We release the replicate package of R.R. at a link.

Be Cautious When Merging Unfamiliar LLMs: A Phishing Model Capable of Stealing Privacy

Model merging is a widespread technology in large language models (LLMs) that integrates multiple task-specific LLMs into a unified one, enabling the merged model to inherit the specialized capabilities of these LLMs. Most task-specific LLMs are sourced from open-source communities and have not undergone rigorous auditing, potentially imposing risks in model merging. This paper highlights an overlooked privacy risk: \textit{an unsafe model could compromise the privacy of other LLMs involved in the model merging.} Specifically, we propose PhiMM, a privacy attack approach that trains a phishing model capable of stealing privacy using a crafted privacy phishing instruction dataset. Furthermore, we introduce a novel model cloaking method that mimics a specialized capability to conceal attack intent, luring users into merging the phishing model. Once victims merge the phishing model, the attacker can extract personally identifiable information (PII) or infer membership information (MI) by querying the merged model with the phishing instruction. Experimental results show that merging a phishing model increases the risk of privacy breaches. Compared to the results before merging, PII leakage increased by 3.9\% and MI leakage increased by 17.4\% on average. We release the code of PhiMM through a link.

Could It Be Generated? Towards Practical Analysis of Memorization in Text-To-Image Diffusion Models

The past few years have witnessed substantial advancement in text-guided image generation powered by diffusion models. However, it was shown that text-to-image diffusion models are vulnerable to training image memorization, raising concerns on copyright infringement and privacy invasion. In this work, we perform practical analysis of memorization in text-to-image diffusion models. Targeting a set of images to protect, we conduct quantitive analysis on them without need to collect any prompts. Specifically, we first formally define the memorization of image and identify three necessary conditions of memorization, respectively similarity, existence and probability. We then reveal the correlation between the model's prediction error and image replication. Based on the correlation, we propose to utilize inversion techniques to verify the safety of target images against memorization and measure the extent to which they are memorized. Model developers can utilize our analysis method to discover memorized images or reliably claim safety against memorization. Extensive experiments on the Stable Diffusion, a popular open-source text-to-image diffusion model, demonstrate the effectiveness of our analysis method.

How ChatGPT is Solving Vulnerability Management Problem

Recently, ChatGPT has attracted great attention from the code analysis domain. Prior works show that ChatGPT has the capabilities of processing foundational code analysis tasks, such as abstract syntax tree generation, which indicates the potential of using ChatGPT to comprehend code syntax and static behaviors. However, it is unclear whether ChatGPT can complete more complicated real-world vulnerability management tasks, such as the prediction of security relevance and patch correctness, which require an all-encompassing understanding of various aspects, including code syntax, program semantics, and related manual comments. In this paper, we explore ChatGPT's capabilities on 6 tasks involving the complete vulnerability management process with a large-scale dataset containing 78,445 samples. For each task, we compare ChatGPT against SOTA approaches, investigate the impact of different prompts, and explore the difficulties. The results suggest promising potential in leveraging ChatGPT to assist vulnerability management. One notable example is ChatGPT's proficiency in tasks like generating titles for software bug reports. Furthermore, our findings reveal the difficulties encountered by ChatGPT and shed light on promising future directions. For instance, directly providing random demonstration examples in the prompt cannot consistently guarantee good performance in vulnerability management. By contrast, leveraging ChatGPT in a self-heuristic way -- extracting expertise from demonstration examples itself and integrating the extracted expertise in the prompt is a promising research direction. Besides, ChatGPT may misunderstand and misuse the information in the prompt. Consequently, effectively guiding ChatGPT to focus on helpful information rather than the irrelevant content is still an open problem.

LMSanitator: Defending Prompt-Tuning Against Task-Agnostic Backdoors

Prompt-tuning has emerged as an attractive paradigm for deploying large-scale language models due to its strong downstream task performance and efficient multitask serving ability. Despite its wide adoption, we empirically show that prompt-tuning is vulnerable to downstream task-agnostic backdoors, which reside in the pretrained models and can affect arbitrary downstream tasks. The state-of-the-art backdoor detection approaches cannot defend against task-agnostic backdoors since they hardly converge in reversing the backdoor triggers. To address this issue, we propose LMSanitator, a novel approach for detecting and removing task-agnostic backdoors on Transformer models. Instead of directly inversing the triggers, LMSanitator aims to inverse the predefined attack vectors (pretrained models' output when the input is embedded with triggers) of the task-agnostic backdoors, which achieves much better convergence performance and backdoor detection accuracy. LMSanitator further leverages prompt-tuning's property of freezing the pretrained model to perform accurate and fast output monitoring and input purging during the inference phase. Extensive experiments on multiple language models and NLP tasks illustrate the effectiveness of LMSanitator. For instance, LMSanitator achieves 92.8% backdoor detection accuracy on 960 models and decreases the attack success rate to less than 1% in most scenarios.