HInter: Exposing Hidden Intersectional Bias in Large Language Models

Large Language Models (LLMs) may portray discrimination towards certain individuals, especially those characterized by multiple attributes (aka intersectional bias). Discovering intersectional bias in LLMs is challenging, as it involves complex inputs on multiple attributes (e.g. race and gender). To address this challenge, we propose HInter, a test technique that synergistically combines mutation analysis, dependency parsing and metamorphic oracles to automatically detect intersectional bias in LLMs. HInter generates test inputs by systematically mutating sentences using multiple mutations, validates inputs via a dependency invariant and detects biases by checking the LLM response on the original and mutated sentences. We evaluate HInter using six LLM architectures and 18 LLM models (GPT3.5, Llama2, BERT, etc) and find that 14.61% of the inputs generated by HInter expose intersectional bias. Results also show that our dependency invariant reduces false positives (incorrect test inputs) by an order of magnitude. Finally, we observed that 16.62% of intersectional bias errors are hidden, meaning that their corresponding atomic cases do not trigger biases. Overall, this work emphasize the importance of testing LLMs for intersectional bias.

Robustness Analysis of AI Models in Critical Energy Systems

This paper analyzes the robustness of state-of-the-art AI-based models for power grid operations under the $N-1$ security criterion. While these models perform well in regular grid settings, our results highlight a significant loss in accuracy following the disconnection of a line.%under this security criterion. Using graph theory-based analysis, we demonstrate the impact of node connectivity on this loss. Our findings emphasize the need for practical scenario considerations in developing AI methodologies for critical infrastructure.

Constrained Adaptive Attacks: Realistic Evaluation of Adversarial Examples and Robust Training of Deep Neural Networks for Tabular Data

State-of-the-art deep learning models for tabular data have recently achieved acceptable performance to be deployed in industrial settings. However, the robustness of these models remains scarcely explored. Contrary to computer vision, there is to date no realistic protocol to properly evaluate the adversarial robustness of deep tabular models due to intrinsic properties of tabular data such as categorical features, immutability, and feature relationship constraints. To fill this gap, we propose CAA, the first efficient evasion attack for constrained tabular deep learning models. CAA is an iterative parameter-free attack that combines gradient and search attacks to generate adversarial examples under constraints. We leverage CAA to build a benchmark of deep tabular models across three popular use cases: credit scoring, phishing and botnet attacks detection. Our benchmark supports ten threat models with increasing capabilities of the attacker, and reflects real-world attack scenarios for each use case. Overall, our results demonstrate how domain knowledge, adversarial training, and attack budgets impact the robustness assessment of deep tabular models and provide security practitioners with a set of recommendations to improve the robustness of deep tabular models against various evasion attack scenarios.

Evaluating the Robustness of Test Selection Methods for Deep Neural Networks

Testing deep learning-based systems is crucial but challenging due to the required time and labor for labeling collected raw data. To alleviate the labeling effort, multiple test selection methods have been proposed where only a subset of test data needs to be labeled while satisfying testing requirements. However, we observe that such methods with reported promising results are only evaluated under simple scenarios, e.g., testing on original test data. This brings a question to us: are they always reliable? In this paper, we explore when and to what extent test selection methods fail for testing. Specifically, first, we identify potential pitfalls of 11 selection methods from top-tier venues based on their construction. Second, we conduct a study on five datasets with two model architectures per dataset to empirically confirm the existence of these pitfalls. Furthermore, we demonstrate how pitfalls can break the reliability of these methods. Concretely, methods for fault detection suffer from test data that are: 1) correctly classified but uncertain, or 2) misclassified but confident. Remarkably, the test relative coverage achieved by such methods drops by up to 86.85%. On the other hand, methods for performance estimation are sensitive to the choice of intermediate-layer output. The effectiveness of such methods can be even worse than random selection when using an inappropriate layer.

CodeLens: An Interactive Tool for Visualizing Code Representations

Representing source code in a generic input format is crucial to automate software engineering tasks, e.g., applying machine learning algorithms to extract information. Visualizing code representations can further enable human experts to gain an intuitive insight into the code. Unfortunately, as of today, there is no universal tool that can simultaneously visualise different types of code representations. In this paper, we introduce a tool, CodeLens, which provides a visual interaction environment that supports various representation methods and helps developers understand and explore them. CodeLens is designed to support multiple programming languages, such as Java, Python, and JavaScript, and four types of code representations, including sequence of tokens, abstract syntax tree (AST), data flow graph (DFG), and control flow graph (CFG). By using CodeLens, developers can quickly visualize the specific code representation and also obtain the represented inputs for models of code. The Web-based interface of CodeLens is available at http://www.codelens.org. The demonstration video can be found at http://www.codelens.org/demo.

RNNS: Representation Nearest Neighbor Search Black-Box Attack on Code Models

Pre-trained code models are mainly evaluated using the in-distribution test data. The robustness of models, i.e., the ability to handle hard unseen data, still lacks evaluation. In this paper, we propose a novel search-based black-box adversarial attack guided by model behaviours for pre-trained programming language models, named Representation Nearest Neighbor Search(RNNS), to evaluate the robustness of Pre-trained PL models. Unlike other black-box adversarial attacks, RNNS uses the model-change signal to guide the search in the space of the variable names collected from real-world projects. Specifically, RNNS contains two main steps, 1) indicate which variable (attack position location) we should attack based on model uncertainty, and 2) search which adversarial tokens we should use for variable renaming according to the model behaviour observations. We evaluate RNNS on 6 code tasks (e.g., clone detection), 3 programming languages (Java, Python, and C), and 3 pre-trained code models: CodeBERT, GraphCodeBERT, and CodeT5. The results demonstrate that RNNS outperforms the state-of-the-art black-box attacking methods (MHM and ALERT) in terms of attack success rate (ASR) and query times (QT). The perturbation of generated adversarial examples from RNNS is smaller than the baselines with respect to the number of replaced variables and the variable length change. Our experiments also show that RNNS is efficient in attacking the defended models and is useful for adversarial training.

Distribution-aware Fairness Test Generation

This work addresses how to validate group fairness in image recognition software. We propose a distribution-aware fairness testing approach (called DistroFair) that systematically exposes class-level fairness violations in image classifiers via a synergistic combination of out-of-distribution (OOD) testing and semantic-preserving image mutation. DistroFair automatically learns the distribution (e.g., number/orientation) of objects in a set of images. Then it systematically mutates objects in the images to become OOD using three semantic-preserving image mutations -- object deletion, object insertion and object rotation. We evaluate DistroFair using two well-known datasets (CityScapes and MS-COCO) and three major, commercial image recognition software (namely, Amazon Rekognition, Google Cloud Vision and Azure Computer Vision). Results show that about 21% of images generated by DistroFair reveal class-level fairness violations using either ground truth or metamorphic oracles. DistroFair is up to 2.3x more effective than two main baselines, i.e., (a) an approach which focuses on generating images only within the distribution (ID) and (b) fairness analysis using only the original image dataset. We further observed that DistroFair is efficient, it generates 460 images per hour, on average. Finally, we evaluate the semantic validity of our approach via a user study with 81 participants, using 30 real images and 30 corresponding mutated images generated by DistroFair. We found that images generated by DistroFair are 80% as realistic as real-world images.

Going Further: Flatness at the Rescue of Early Stopping for Adversarial Example Transferability

Transferability is the property of adversarial examples to be misclassified by other models than the surrogate model for which they were crafted. Previous research has shown that transferability is substantially increased when the training of the surrogate model has been early stopped. A common hypothesis to explain this is that the later training epochs are when models learn the non-robust features that adversarial attacks exploit. Hence, an early stopped model is more robust (hence, a better surrogate) than fully trained models. We demonstrate that the reasons why early stopping improves transferability lie in the side effects it has on the learning dynamics of the model. We first show that early stopping benefits transferability even on models learning from data with non-robust features. We then establish links between transferability and the exploration of the loss landscape in the parameter space, on which early stopping has an inherent effect. More precisely, we observe that transferability peaks when the learning rate decays, which is also the time at which the sharpness of the loss significantly drops. This leads us to propose RFN, a new approach for transferability that minimizes loss sharpness during training in order to maximize transferability. We show that by searching for large flat neighborhoods, RFN always improves over early stopping (by up to 47 points of transferability rate) and is competitive to (if not better than) strong state-of-the-art baselines.

Boosting Source Code Learning with Data Augmentation: An Empirical Study

The next era of program understanding is being propelled by the use of machine learning to solve software problems. Recent studies have shown surprising results of source code learning, which applies deep neural networks (DNNs) to various critical software tasks, e.g., bug detection and clone detection. This success can be greatly attributed to the utilization of massive high-quality training data, and in practice, data augmentation, which is a technique used to produce additional training data, has been widely adopted in various domains, such as computer vision. However, in source code learning, data augmentation has not been extensively studied, and existing practice is limited to simple syntax-preserved methods, such as code refactoring. Essentially, source code is often represented in two ways, namely, sequentially as text data and structurally as graph data, when it is used as training data in source code learning. Inspired by these analogy relations, we take an early step to investigate whether data augmentation methods that are originally used for text and graphs are effective in improving the training quality of source code learning. To that end, we first collect and categorize data augmentation methods in the literature. Second, we conduct a comprehensive empirical study on four critical tasks and 11 DNN architectures to explore the effectiveness of 12 data augmentation methods (including code refactoring and 11 other methods for text and graph data). Our results identify the data augmentation methods that can produce more accurate and robust models for source code learning, including those based on mixup (e.g., SenMixup for texts and Manifold-Mixup for graphs), and those that slightly break the syntax of source code (e.g., random swap and random deletion for texts).

GAT: Guided Adversarial Training with Pareto-optimal Auxiliary Tasks

While leveraging additional training data is well established to improve adversarial robustness, it incurs the unavoidable cost of data collection and the heavy computation to train models. To mitigate the costs, we propose \textit{Guided Adversarial Training } (GAT), a novel adversarial training technique that exploits auxiliary tasks under a limited set of training data. Our approach extends single-task models into multi-task models during the min-max optimization of adversarial training, and drives the loss optimization with a regularization of the gradient curvature across multiple tasks. GAT leverages two types of auxiliary tasks: self-supervised tasks, where the labels are generated automatically, and domain-knowledge tasks, where human experts provide additional labels. Experimentally, under limited data, GAT increases the robust accuracy on CIFAR-10 up to four times (from 11% to 42% robust accuracy) and the robust AUC of CheXpert medical imaging dataset from 50\% to 83\%. On the full CIFAR-10 dataset, GAT outperforms eight state-of-the-art adversarial training strategies. Our large study across five datasets and six tasks demonstrates that task augmentation is an efficient alternative to data augmentation, and can be key to achieving both clean and robust performances.