Generating accurate step-by-step reasoning is essential for Large Language Models (LLMs) to address complex problems and enhance robustness and interpretability. Despite the flux of research on developing advanced reasoning approaches, systematically analyzing the diverse LLMs and reasoning strategies in generating reasoning chains remains a significant challenge. The difficulties stem from the lack of two key elements: (1) an automatic method for evaluating the generated reasoning chains on different tasks, and (2) a unified formalism and implementation of the diverse reasoning approaches for systematic comparison. This paper aims to close the gap: (1) We introduce AutoRace for fully automated reasoning chain evaluation. Existing metrics rely on expensive human annotations or pre-defined LLM prompts not adaptable to different tasks. In contrast, AutoRace automatically creates detailed evaluation criteria tailored for each task, and uses GPT-4 for accurate evaluation following the criteria. (2) We develop LLM Reasoners, a library for standardized modular implementation of existing and new reasoning algorithms, under a unified formulation of the search, reward, and world model components. With the new evaluation and library, (3) we conduct extensive study of different reasoning approaches (e.g., CoT, ToT, RAP). The analysis reveals interesting findings about different factors contributing to reasoning, including the reward-guidance, breadth-vs-depth in search, world model, and prompt formats, etc.

Recent advances in the area of long document matching have primarily focused on using transformer-based models for long document encoding and matching. There are two primary challenges associated with these models. Firstly, the performance gain provided by transformer-based models comes at a steep cost - both in terms of the required training time and the resource (memory and energy) consumption. The second major limitation is their inability to handle more than a pre-defined input token length at a time. In this work, we empirically demonstrate the effectiveness of simple neural models (such as feed-forward networks, and CNNs) and simple embeddings (like GloVe, and Paragraph Vector) over transformer-based models on the task of document matching. We show that simple models outperform the more complex BERT-based models while taking significantly less training time, energy, and memory. The simple models are also more robust to variations in document length and text perturbations.

Recent advancements in deep learning techniques have transformed the area of semantic text matching (STM). However, most state-of-the-art models are designed to operate with short documents such as tweets, user reviews, comments, and so on. These models have fundamental limitations when applied to long-form documents such as scientific papers, legal documents, and patents. When handling such long documents, there are three primary challenges: (i) the presence of different contexts for the same word throughout the document, (ii) small sections of contextually similar text between two documents, but dissimilar text in the remaining parts (this defies the basic understanding of “similarity”), and (iii) the coarse nature of a single global similarity measure which fails to capture the heterogeneity of the document content. In this article, we describe CoLDE : Co ntrastive L ong D ocument E ncoder—a transformer-based framework that addresses these challenges and allows for interpretable comparisons of long documents. CoLDE uses unique positional embeddings and a multi-headed chunkwise attention layer in conjunction with a supervised contrastive learning framework to capture similarity at three different levels: (i) high-level similarity scores between a pair of documents, (ii) similarity scores between different sections within and across documents, and (iii) similarity scores between different chunks in the same document and across other documents. These fine-grained similarity scores aid in better interpretability. We evaluate CoLDE on three long document datasets namely, ACL Anthology publications, Wikipedia articles, and USPTO patents. Besides outperforming the state-of-the-art methods on the document matching task, CoLDE is also robust to changes in document length and text perturbations and provides interpretable results. The code for the proposed model is publicly available at https://github.com/InterDigitalInc/CoLDE .

Deep neural classifiers trained with cross-entropy loss (CE loss) often suffer from poor calibration, necessitating the task of out-of-distribution (OOD) detection. Traditional supervised OOD detection methods require expensive manual annotation of in-distribution and OOD samples. To address the annotation bottleneck, we introduce SELFOOD, a self-supervised OOD detection method that requires only in-distribution samples as supervision. We cast OOD detection as an inter-document intra-label (IDIL) ranking problem and train the classifier with our pairwise ranking loss, referred to as IDIL loss. Specifically, given a set of in-distribution documents and their labels, for each label, we train the classifier to rank the softmax scores of documents belonging to that label to be higher than the scores of documents that belong to other labels. Unlike CE loss, our IDIL loss function reaches zero when the desired confidence ranking is achieved and gradients are backpropagated to decrease probabilities associated with incorrect labels rather than continuously increasing the probability of the correct label. Extensive experiments with several classifiers on multiple classification datasets demonstrate the effectiveness of our method in both coarse- and fine-grained settings.

The ever-increasing number of android malware still poses a critical security challenge to the smartphone ecosystem. Literature suggests that machine and deep learning models can detect android malware with high accuracy and low false positivity. However, the result of arms-race in the adversarial setting of these detection models will shape their integration in real-world applications. Therefore, we first constructed four different malware detection models by applying machine and deep learning algorithms. Then, we stepped into the malware developer's shoes and created an adversarial setting framework to investigate the robustness of the above detection models. We developed an evasion attack (Gradient Modification Attack) to exploit the vulnerabilities and force massive misclassifications in the above detection models. The attack drastically reduces the average accuracy of the above four detection models from 95.13% to 59.97%. Later, we also developed a potential defense mechanism (Correlated Distillation Retraining) to mitigate such adversarial attacks. In the end, we conclude that investigation of malware detection models in adversarial settings is essential for improving their robustness and real-world deployment.

The android ecosystem (smartphones, tablets, etc.) has grown manifold in the last decade. However, the exponential surge of android malware is threatening the ecosystem. Literature suggests that android malware can be detected using machine and deep learning classifiers; however, these detection models might be vulnerable to adversarial attacks. This work investigates the adversarial robustness of twenty-four diverse malware detection models developed using two features and twelve learning algorithms across four categories (machine learning, bagging classifiers, boosting classifiers, and neural network). We stepped into the adversary’s shoes and proposed two false-negative evasion attacks, namely GradAA and GreedAA , to expose vulnerabilities in the above detection models. The evasion attack agents transform malware applications into adversarial malware applications by adding minimum noise (maximum five perturbations) while maintaining the modified applications’ structural, syntactic, and behavioral integrity. These adversarial malware applications force misclassifications and are predicted as benign by the detection models. The evasion attacks achieved an average fooling rate of 83.34 % (GradAA) and 99.21 % (GreedAA) which reduced the average accuracy from 90.35 % to 55.22 % (GradAA) and 48.29 % (GreedAA) in twenty-four detection models. We also proposed two defense strategies, namely Adversarial Retraining and Correlation Distillation Retraining as countermeasures to protect detection models from adversarial attacks. The defense strategies slightly improved the detection accuracy but drastically enhanced the adversarial robustness of detection models. Finally, investigating the robustness of malware detection models against adversarial attacks is an essential step before their real-world deployment and can help in developing adversarially superior detection models.

Research shows that recent information spreading on social media is dubious and untrue, intended to mislead audiences. In the last few years, the fake news spreading community has emerged on a large scale, which intentionally promotes incorrect information, becoming challenging to detect even with human annotation. This paper aims to profile fake news and investigate the characteristics that make the news untrue and fake. We validate our features set on four open-source datasets popularly used in the community. We employ various machine learning, artificial, and recurrent neural network models to evaluate the proposed features set’s performance and importance. Our results show that proposed features are generalized and not specific to a domain. Our findings reveal that the model outperforms raw content-based datasets (ISOT and Fake News Master) than interpretation-based datasets (Liar-Liar and Kaggle Emergent).

The last decade witnessed an exponential growth of smartphones and their users, which has drawn massive attention from malware designers. The current malware detection engines are unable to cope with the volume, velocity, and variety of incoming malware. Thus the anti-malware community is investigating the use of machine learning and deep learning to develop malware detection models. However, research in other domains suggests that the machine learning/deep learning models are vulnerable to adversarial attacks. Therefore in this work, we proposed a framework to construct robust malware detection models against adversarial attacks. We first constructed twelve different malware detection models using a variety of classification algorithms. Then we acted as an adversary and proposed Gradient-based Adversarial Attack Network to perform adversarial attacks on the above detection models. The attack is designed to convert the maximum number of malware samples into adversarial samples with minimal modifications in each sample. The proposed attack achieves an average fooling rate of 98.68% against twelve permission-based malware detection models and 90.71% against twelve intent-based malware detection models. We also identified the list of vulnerable permissions/intents which an adversary can use to force misclassifications in detection models. Later we proposed three adversarial defense strategies to counter the attacks performed on detection models. The proposed Hybrid Distillation based defense strategy improved the average accuracy by 54.21% for twelve permission-based detection models and 59.14% for intent-based detection models. We also concluded that the adversarial-based study improves the performance and robustness of malware detection models and is essential before any real-world deployment.