TrustLLM: Trustworthiness in Large Language Models

Large language models (LLMs), exemplified by ChatGPT, have gained considerable attention for their excellent natural language processing capabilities. Nonetheless, these LLMs present many challenges, particularly in the realm of trustworthiness. Therefore, ensuring the trustworthiness of LLMs emerges as an important topic. This paper introduces TrustLLM, a comprehensive study of trustworthiness in LLMs, including principles for different dimensions of trustworthiness, established benchmark, evaluation, and analysis of trustworthiness for mainstream LLMs, and discussion of open challenges and future directions. Specifically, we first propose a set of principles for trustworthy LLMs that span eight different dimensions. Based on these principles, we further establish a benchmark across six dimensions including truthfulness, safety, fairness, robustness, privacy, and machine ethics. We then present a study evaluating 16 mainstream LLMs in TrustLLM, consisting of over 30 datasets. Our findings firstly show that in general trustworthiness and utility (i.e., functional effectiveness) are positively related. Secondly, our observations reveal that proprietary LLMs generally outperform most open-source counterparts in terms of trustworthiness, raising concerns about the potential risks of widely accessible open-source LLMs. However, a few open-source LLMs come very close to proprietary ones. Thirdly, it is important to note that some LLMs may be overly calibrated towards exhibiting trustworthiness, to the extent that they compromise their utility by mistakenly treating benign prompts as harmful and consequently not responding. Finally, we emphasize the importance of ensuring transparency not only in the models themselves but also in the technologies that underpin trustworthiness. Knowing the specific trustworthy technologies that have been employed is crucial for analyzing their effectiveness.

SneakyPrompt: Evaluating Robustness of Text-to-image Generative Models' Safety Filters

Text-to-image generative models such as Stable Diffusion and DALL$\cdot$E 2 have attracted much attention since their publication due to their wide application in the real world. One challenging problem of text-to-image generative models is the generation of Not-Safe-for-Work (NSFW) content, e.g., those related to violence and adult. Therefore, a common practice is to deploy a so-called safety filter, which blocks NSFW content based on either text or image features. Prior works have studied the possible bypass of such safety filters. However, existing works are largely manual and specific to Stable Diffusion's official safety filter. Moreover, the bypass ratio of Stable Diffusion's safety filter is as low as 23.51% based on our evaluation. In this paper, we propose the first automated attack framework, called SneakyPrompt, to evaluate the robustness of real-world safety filters in state-of-the-art text-to-image generative models. Our key insight is to search for alternative tokens in a prompt that generates NSFW images so that the generated prompt (called an adversarial prompt) bypasses existing safety filters. Specifically, SneakyPrompt utilizes reinforcement learning (RL) to guide an agent with positive rewards on semantic similarity and bypass success. Our evaluation shows that SneakyPrompt successfully generated NSFW content using an online model DALL$\cdot$E 2 with its default, closed-box safety filter enabled. At the same time, we also deploy several open-source state-of-the-art safety filters on a Stable Diffusion model and show that SneakyPrompt not only successfully generates NSFW content, but also outperforms existing adversarial attacks in terms of the number of queries and image qualities.

Addressing Heterogeneity in Federated Learning via Distributional Transformation

Federated learning (FL) allows multiple clients to collaboratively train a deep learning model. One major challenge of FL is when data distribution is heterogeneous, i.e., differs from one client to another. Existing personalized FL algorithms are only applicable to narrow cases, e.g., one or two data classes per client, and therefore they do not satisfactorily address FL under varying levels of data heterogeneity. In this paper, we propose a novel framework, called DisTrans, to improve FL performance (i.e., model accuracy) via train and test-time distributional transformations along with a double-input-channel model structure. DisTrans works by optimizing distributional offsets and models for each FL client to shift their data distribution, and aggregates these offsets at the FL server to further improve performance in case of distributional heterogeneity. Our evaluation on multiple benchmark datasets shows that DisTrans outperforms state-of-the-art FL methods and data augmentation methods under various settings and different degrees of client distributional heterogeneity.

EdgeMixup: Improving Fairness for Skin Disease Classification and Segmentation

Skin lesions can be an early indicator of a wide range of infectious and other diseases. The use of deep learning (DL) models to diagnose skin lesions has great potential in assisting clinicians with prescreening patients. However, these models often learn biases inherent in training data, which can lead to a performance gap in the diagnosis of people with light and/or dark skin tones. To the best of our knowledge, limited work has been done on identifying, let alone reducing, model bias in skin disease classification and segmentation. In this paper, we examine DL fairness and demonstrate the existence of bias in classification and segmentation models for subpopulations with darker skin tones compared to individuals with lighter skin tones, for specific diseases including Lyme, Tinea Corporis and Herpes Zoster. Then, we propose a novel preprocessing, data alteration method, called EdgeMixup, to improve model fairness with a linear combination of an input skin lesion image and a corresponding a predicted edge detection mask combined with color saturation alteration. For the task of skin disease classification, EdgeMixup outperforms much more complex competing methods such as adversarial approaches, achieving a 10.99% reduction in accuracy gap between light and dark skin tone samples, and resulting in 8.4% improved performance for an underrepresented subpopulation.

Defending Medical Image Diagnostics against Privacy Attacks using Generative Methods

Machine learning (ML) models used in medical imaging diagnostics can be vulnerable to a variety of privacy attacks, including membership inference attacks, that lead to violations of regulations governing the use of medical data and threaten to compromise their effective deployment in the clinic. In contrast to most recent work in privacy-aware ML that has been focused on model alteration and post-processing steps, we propose here a novel and complementary scheme that enhances the security of medical data by controlling the data sharing process. We develop and evaluate a privacy defense protocol based on using a generative adversarial network (GAN) that allows a medical data sourcer (e.g. a hospital) to provide an external agent (a modeler) a proxy dataset synthesized from the original images, so that the resulting diagnostic systems made available to model consumers is rendered resilient to privacy attackers. We validate the proposed method on retinal diagnostics AI used for diabetic retinopathy that bears the risk of possibly leaking private information. To incorporate concerns of both privacy advocates and modelers, we introduce a metric to evaluate privacy and utility performance in combination, and demonstrate, using these novel and classical metrics, that our approach, by itself or in conjunction with other defenses, provides state of the art (SOTA) performance for defending against privacy attacks.

Practical Blind Membership Inference Attack via Differential Comparisons

Membership inference (MI) attacks affect user privacy by inferring whether given data samples have been used to train a target learning model, e.g., a deep neural network. There are two types of MI attacks in the literature, i.e., these with and without shadow models. The success of the former heavily depends on the quality of the shadow model, i.e., the transferability between the shadow and the target; the latter, given only blackbox probing access to the target model, cannot make an effective inference of unknowns, compared with MI attacks using shadow models, due to the insufficient number of qualified samples labeled with ground truth membership information. In this paper, we propose an MI attack, called BlindMI, which probes the target model and extracts membership semantics via a novel approach, called differential comparison. The high-level idea is that BlindMI first generates a dataset with nonmembers via transforming existing samples into new samples, and then differentially moves samples from a target dataset to the generated, non-member set in an iterative manner. If the differential move of a sample increases the set distance, BlindMI considers the sample as non-member and vice versa. BlindMI was evaluated by comparing it with state-of-the-art MI attack algorithms. Our evaluation shows that BlindMI improves F1-score by nearly 20% when compared to state-of-the-art on some datasets, such as Purchase-50 and Birds-200, in the blind setting where the adversary does not know the target model's architecture and the target dataset's ground truth labels. We also show that BlindMI can defeat state-of-the-art defenses.

PatchAttack: A Black-box Texture-based Attack with Reinforcement Learning

Patch-based attacks introduce a perceptible but localized change to the input that induces misclassification. A limitation of current patch-based black-box attacks is that they perform poorly for targeted attacks, and even for the less challenging non-targeted scenarios, they require a large number of queries. Our proposed PatchAttack is query efficient and can break models for both targeted and non-targeted attacks. PatchAttack induces misclassifications by superimposing small textured patches on the input image. We parametrize the appearance of these patches by a dictionary of class-specific textures. This texture dictionary is learned by clustering Gram matrices of feature activations from a VGG backbone. PatchAttack optimizes the position and texture parameters of each patch using reinforcement learning. Our experiments show that PatchAttack achieves > 99% success rate on ImageNet for a wide range of architectures, while only manipulating 3% of the image for non-targeted attacks and 10% on average for targeted attacks. Furthermore, we show that PatchAttack circumvents state-of-the-art adversarial defense methods successfully.

Towards Practical Verification of Machine Learning: The Case of Computer Vision Systems

Due to the increasing usage of machine learning (ML) techniques in security- and safety-critical domains, such as autonomous systems and medical diagnosis, ensuring correct behavior of ML systems, especially for different corner cases, is of growing importance. In this paper, we propose a generic framework for evaluating security and robustness of ML systems using different real-world safety properties. We further design, implement and evaluate VeriVis, a scalable methodology that can verify a diverse set of safety properties for state-of-the-art computer vision systems with only blackbox access. VeriVis leverage different input space reduction techniques for efficient verification of different safety properties. VeriVis is able to find thousands of safety violations in fifteen state-of-the-art computer vision systems including ten Deep Neural Networks (DNNs) such as Inception-v3 and Nvidia's Dave self-driving system with thousands of neurons as well as five commercial third-party vision APIs including Google vision and Clarifai for twelve different safety properties. Furthermore, VeriVis can successfully verify local safety properties, on average, for around 31.7% of the test images. VeriVis finds up to 64.8x more violations than existing gradient-based methods that, unlike VeriVis, cannot ensure non-existence of any violations. Finally, we show that retraining using the safety violations detected by VeriVis can reduce the average number of violations up to 60.2%.