Large Language Models (LLMs), pre-trained on massive text corpora, exhibit remarkable human-level language understanding, reasoning, and decision-making abilities. However, they tend to memorize unwanted information, such as private or copyrighted content, raising significant privacy and legal concerns. Unlearning has emerged as a promising solution, but existing methods face a significant challenge of over-forgetting. This issue arises because they indiscriminately suppress the generation of all the tokens in forget samples, leading to a substantial loss of model utility. To overcome this challenge, we introduce the Targeted Information Forgetting (TIF) framework, which consists of (1) a flexible targeted information identifier designed to differentiate between unwanted words (UW) and general words (GW) in the forget samples, and (2) a novel Targeted Preference Optimization approach that leverages Logit Preference Loss to unlearn unwanted information associated with UW and Preservation Loss to retain general information in GW, effectively improving the unlearning process while mitigating utility degradation. Extensive experiments on the TOFU and MUSE benchmarks demonstrate that the proposed TIF framework enhances unlearning effectiveness while preserving model utility and achieving state-of-the-art results.

Membership Inference Attacks (MIAs) have recently been employed to determine whether a specific text was part of the pre-training data of Large Language Models (LLMs). However, existing methods often misinfer non-members as members, leading to a high false positive rate, or depend on additional reference models for probability calibration, which limits their practicality. To overcome these challenges, we introduce a novel framework called Automatic Calibration Membership Inference Attack (ACMIA), which utilizes a tunable temperature to calibrate output probabilities effectively. This approach is inspired by our theoretical insights into maximum likelihood estimation during the pre-training of LLMs. We introduce ACMIA in three configurations designed to accommodate different levels of model access and increase the probability gap between members and non-members, improving the reliability and robustness of membership inference. Extensive experiments on various open-source LLMs demonstrate that our proposed attack is highly effective, robust, and generalizable, surpassing state-of-the-art baselines across three widely used benchmarks. The source code is publicly available.

Vision Transformers (ViTs) have become prominent models for solving various vision tasks. However, the interpretability of ViTs has not kept pace with their promising performance. While there has been a surge of interest in developing post hoc solutions to explain ViTs’ outputs, these methods do not generalize to different downstream tasks and various transformer architectures. Furthermore, if ViTs are not properly trained with the given data and do not prioritize the region of interest, the post hoc methods become less effective. To overcome this limitation, we introduce a novel training procedure that inherently enhances ViT’s interpretability. Our interpretability-aware ViT (IA-ViT) draws inspiration from a fresh insight: both the class patch and image patches consistently generate predicted distributions and attention maps. IA-ViT is composed of a feature extractor, a predictor, and an interpreter, which are trained jointly with an interpretability-aware training objective. Consequently, the interpreter simulates the behavior of the predictor and provides a faithful explanation through its single-head self-attention mechanism. Our comprehensive experimental results demonstrate the effectiveness of IA-ViT in several image classification tasks, with both qualitative and quantitative evaluations of model performance and interpretability. Our code is available at: https://github.com/qiangyao1988/IA-ViT.

Recent studies have demonstrated significant success in detecting attacks on the Controller Area Network (CAN) bus network using machine learning and deep learning models, including convolutional neural networks and transformer-based architectures. Building on this foundation, our work investigates the use of large language models (LLMs) not only for intrusion detection but also for providing interpretable explanations of their decisions. We fine-tuned three LLMs, i.e., SecureBERT, LLaMA-2, and LLaMA-3, for intrusion detection on CAN bus data. Among them, LLaMA-3 delivered the best results, achieving SOTA performance on the Car-Hacking dataset. Beyond attack classification, we evaluated LLaMA-3’s ability to generate reasoning for its decisions through zero-shot prompting. The model successfully articulated its rationale, particularly for Denial-of-Service (DoS) attacks, demonstrating strong potential for explain-ability in intrusion detection systems. These findings highlight the potential of LLMs to serve as a highly accurate intrusion detection system while simultaneously providing interpretable explanations, thereby enhancing the investigative capabilities of cybersecurity professionals.

Segment Anything Model (SAM) is one of the pioneering prompt-based foundation models for image segmentation and has been rapidly adopted for various medical imaging applications. However, in clinical settings, cre-ating effective prompts is notably challenging and time-consuming, requiring the expertise of domain specialists such as physicians. This requirement significantly dimin-ishes SAM's primary advantage—its interactive capability with end users—in medical applications. Moreover, recent studies have indicated that SAM, originally designed for 2D natural images, performs suboptimally on 3D medical image segmentation tasks. This subpar performance is attributed to the domain gaps between natural and medical images and the disparities in spatial arrangements between 2D and 3D images, particularly in multi-organ segmentation applications. To overcome these challenges, we present a novel technique termed AutoProSAM. This method au-tomates 3D multi-organ CT-based segmentation by lever-aging SAM's foundational model capabilities without relying on domain experts for prompts. The approach utilizes parameter-efficient adaptation techniques to adapt SAMfor 3D medical imagery and incorporates an effective automatic prompt learning paradigm specific to this domain. By eliminating the need for manual prompts, it enhances SAM's capabilities for 3D medical image segmentation and achieves state-of-the-art (SOTA) performance in CT-based multi-organ segmentation tasks. The code is in this link.

This paper introduces a novel information retrieval (IR) task of Conversational Entity Retrieval from a Knowledge Graph (CER-KG), which extends non-conversational entity retrieval from a knowledge graph (KG) to the conversational scenario. The user queries in CER-KG dialog turns may rely on the results of the preceding turns, which are KG entities. Similar to the conversational document IR, CER-KG can be viewed as a sequence of interrelated ranking tasks. To enable future research on CER-KG, we created QBLink-KG, a publicly available benchmark that was adapted from QBLink, a benchmark for text-based conversational reading comprehension of Wikipedia. As an initial approach to CER-KG, we experimented with Transformer- and LSTM-based query encoders in combination with the Neural Architecture for Conversational Entity Retrieval (NACER), our proposed feature-based neural architecture for entity ranking in CER-KG. NACER computes the ranking score of a candidate KG entity by taking into account diverse lexical and semantic matching signals between various KG components in its neighborhood, such as entities, categories, and literals, as well as entities in the results of the preceding turns in dialog history. The reported experimental results reveal the key challenges of CER-KG along with the possible directions for new approaches to this task.

Deep neural networks (DNNs) for supervised learning can be viewed as a pipeline of the feature extractor (i.e., last hidden layer) and a linear classifier (i.e., output layer) that are trained jointly with stochastic gradient descent (SGD) on the loss function (e.g., cross-entropy). In each epoch, the true gradient of the loss function is estimated using a mini-batch sampled from the training set and model parameters are then updated with the mini-batch gradients. Although the latter provides an unbiased estimation of the former, they are subject to substantial variances derived from the size and number of sampled mini-batches, leading to noisy and jumpy updates. To stabilize such undesirable variance in estimating the true gradients, we propose In-Training Representation Alignment (ITRA) that explicitly aligns feature distributions of two different mini-batches with a matching loss in the SGD training process. We also provide a rigorous analysis of the desirable effects of the matching loss on feature representation learning: (1) extracting compact feature representation; (2) reducing over-adaption on mini-batches via an adaptively weighting mechanism; and (3) accommodating to multi-modalities. Finally, we conduct large-scale experiments on both image and text classifications to demonstrate its superior performance to the strong baselines.

There are increasing demands for understanding deep neural networks' (DNNs) behavior spurred by growing security and/or transparency concerns. Due to multi-layer nonlinearity of the deep neural network architectures, explaining DNN predictions still remains as an open problem, preventing us from gaining a deeper understanding of the mechanisms. To enhance the explainability of DNNs, we estimate the input feature's attributions to the prediction task using divergence and flux. Inspired by the divergence theorem in vector analysis, we develop a novel Negative Flux Aggregation (NeFLAG) formulation and an efficient approximation algorithm to estimate attribution map. Unlike the previous techniques, ours doesn't rely on fitting a surrogate model nor need any path integration of gradients. Both qualitative and quantitative experiments demonstrate a superior performance of NeFLAG in generating more faithful attribution maps than the competing methods. Our code is available at https://github.com/xinli0928/NeFLAG.

Bias in the training data can jeopardize fairness and explainability of deep neural network prediction on test data. We propose a novel bias-tailored data augmentation approach, Counterfactual Interpolation Augmentation (CIA), attempting to debias the training data by d-separating the spurious correlation between the target variable and the sensitive attribute. CIA generates counterfactual interpolations along a path simulating the distribution transitions between the input and its counterfactual example. CIA as a pre-processing approach enjoys two advantages: First, it couples with either plain training or debiasing training to markedly increase fairness over the sensitive attribute. Second, it enhances the explainability of deep neural networks by generating attribution maps via integrating counterfactual gradients. We demonstrate the superior performance of the CIA-trained deep neural network models using qualitative and quantitative experimental results. Our code is available at: https://github.com/qiangyao1988/CIA

Mobile artificial intelligence has recently gained more attention due to the increasing computing power of mobile devices and applications in computer vision, natural language processing, and internet of things. Although large pre-trained language models (e.g., BERT, GPT) have recently achieved the state-of-the-art results on text classification tasks, they are not well suited for latency critical applications on mobile devices. Therefore, it is essential to design tiny models to reduce their memory and computing requirements. Model compression has shown promising results for this goal. However, some significant challenges are yet to be addressed, such as information loss and adversarial robustness. This paper attempts to tackle these challenges through a new training scheme that minimizes the information loss by maximizing the mutual information between the feature representations learned from the large and tiny models. In addition, we propose a certifiably robust defense method named GradMASK that masks a certain proportion of words in an input text. It can defend against both character-level perturbations and word substitution-based attacks. We perform extensive experiments demonstrating the effectiveness of our approach by comparing our tiny RNN models with compact RNNs (e.g., FastGRNN) and compressed RNNs (e.g., PRADO) in clean and adversarial test settings.

Transformers have improved the state-of-the-art in various natural language processing and computer vision tasks. However, the success of the Transformer model has not yet been duly explained. Current explanation techniques, which dissect either the self-attention mechanism or gradient-based attribution, do not necessarily provide a faithful explanation of the inner workings of Transformers due to the following reasons: first, attention weights alone without considering the magnitudes of feature values are not adequate to reveal the self-attention mechanism; second, whereas most Transformer explanation techniques utilize self-attention module, the skip-connection module, contributing a significant portion of information flows in Transformers, has not yet been sufficiently exploited in explanation; third, the gradient-based attribution of individual feature does not incorporate interaction among features in explaining the model's output. In order to tackle the above problems, we propose a novel Transformer explanation technique via attentive class activation tokens, aka, AttCAT, leveraging encoded features, their gradients, and their attention weights to generate a faithful and confident explanation for Transformer's output. Extensive experiments are conducted to demonstrate the superior performance of AttCAT, which generalizes well to different Transformer architectures, evaluation metrics, datasets, and tasks, to the baseline methods. Our code is available at: https://github.com/qiangyao1988/AttCAT.

Existing aspect based sentiment analysis (ABSA) approaches leverage various neural network models to extract the aspect sentiments via learning aspect-specific feature representations. However, these approaches heavily rely on manual tagging of user reviews according to the predefined aspects as the input, a laborious and time-consuming process. Moreover, the underlying methods do not explain how and why the opposing aspect level polarities in a user review lead to the overall polarity. In this paper, we tackle these two problems by designing and implementing a new Multiple-Attention Network (MAN) approach for more powerful ABSA without the need for aspect tags using two new tag-free data sets crawled directly from TripAdvisor (https://www.tripadvisor.com). With the Self- and Position-Aware attention mechanism, MAN is capable of extracting both aspect level and overall sentiments from the text reviews using the aspect level and overall customer ratings, and it can also detect the vital aspect(s) leading to the overall sentiment polarity among different aspects via a new aspect ranking scheme. We carry out extensive experiments to demonstrate the strong performance of MAN compared to other state-of-the-art ABSA approaches and the explainability of our approach by visualizing and interpreting attention weights in case studies.