Abstract A fundamental limitation of modern conversational AI is its limited capacity to demonstrate sustained empathy in long-form interactions. We propose SCIRAG (Semantic Context Improvisational Retrieval-Augmented Generation), a feedback-driven retrieval framework for adaptive empathic dialogue. It employs a dual-loop retrieval framework, iteratively optimizing a static counseling dataset through user metadata and feedback memory refinement. To enhance contextual alignment, we deploy retrieval adaptation, enabling the model to retain and leverage past conversational cues based on user preferences. When integrated with Mixtral-8x7B, SCIRAG improves human-rated empathic understanding by +1.26 points and empathic response by +1.00 point on the RoPE scale, while increasing acceptability by +7.66 points compared to a fine-tuned non-RAG baseline. Automatic evaluation further shows gains in semantic alignment (BERTScore-F1 +0.11) and fluency (perplexity reduced from 19.1 to 12.3). We also present EMPATHIC , a dataset of unscripted, therapeutic conversations. Unlike conventional datasets that contain only 2-4 dialogue turns per conversation, EMPATHIC provides extended conversational trajectories (50+ turns per session), allowing models to learn long-range coherence and empathic listening. The proposed dataset will be publicly released.

Out-of-distribution (OOD) generalization is critical for reliable intrusion detection systems (IDS), yet current methods often falter against stealthy, on-manifold adversarial attacks that mimic ID data. To solve this challenge, we propose a semi-supervised approach that applies an invertible transformation to the latent space and leverages changes in differential entropy to detect OOD samples. Experiments on the KDD99 and X-IIoTID datasets demonstrate that our approach outperforms state-of-the-art defenses, providing enhanced robustness and generalizability for IDS.

Network Intrusion Detection Systems (NIDS), which play a crucial role in defending Industrial Internet of Things (IIoT) networks, often utilize Deep Neural Networks (DNN) for their pattern recognition capabilities. However, these systems remain susceptible to sophisticated adversarial attacks, particularly on-manifold and off-manifold attacks, which skillfully evade detection. This paper addresses the limitations in existing research, focusing primarily on: the predominant focus on off-manifold attacks, while often overlooking subtler yet potent on-manifold attacks; a lack of consideration for the functional behavior of these attacks; reliance on detailed knowledge of the target NIDS for creating attacks; and the need for detailed knowledge about the creation process of adversarial attacks for effective detection. This paper introduces the Saliency Adversarial Autoencoder (SAAE), designed for generating on-manifold attacks through latent space perturbations. This dual-space perturbation approach enables SAAE to efficiently create stealthy attacks that blend with normal network behavior, posing significant challenges to state-of-the-art (SOTA) NIDS. To counter these advanced threats, we propose an attack-agnostic defence mechanism utilizing a fusion-based Autoencoder (AE) with disentangled representations. This defence is adept at detecting threats within the manifold, significantly enhancing NIDS robustness. Comparative assessments with SOTA DNN and Deep Reinforcement Learning (DRL) models highlight the effectiveness of our approach. The SAAE model markedly reduces True Positive Rates (TPR) in these systems. For DNNs, TPR dropped from 99.72% to 41.5%, and for DRLs, from 95.6% to 63.94%. Conversely, our defence model shows high TPR in detecting these attacks, registering 94% for DNNs and 92% for DRLs. Additionally, we release our dataset, named OOM-X-IIoTID 1 1 The datasets can be found at the following link: https://github.com/mohdah200/OOM-X-IIoTID . , which includes On/Off manifold adversarial attacks, a first in the field, to facilitate further research and development in cybersecurity.

Embodied AI aims to develop robots that can understand and execute human language instructions, as well as communicate in natural languages. On this front, we study the task of generating highly detailed navigational instructions for the embodied robots to follow. Although recent studies have demonstrated significant leaps in the generation of step-by-step instructions from sequences of images, the generated instructions lack variety in terms of their referral to objects and landmarks. Existing speaker models learn strategies to evade the evaluation metrics and obtain higher scores even for low-quality sentences. In this work, we propose SAS (Spatially-Aware Speaker), an instruction generator or Speaker model that utilises both structural and semantic knowledge of the environment to produce richer instructions. For training, we employ a reward learning method in an adversarial setting to avoid systematic bias introduced by language evaluation metrics. Empirically, our method outperforms existing instruction generation models, evaluated using standard metrics. Our code is available at https://github.com/gmuraleekrishna/SAS.

Multi-class multi-instance segmentation is the task of identifying masks for multiple object classes and multiple instances of the same class within an image. The foundational Segment Anything Model (SAM) is designed for promptable multi-class multi-instance segmentation but tends to output part or sub-part masks in the “everything” mode for various real-world applications. Whole object segmentation masks play a crucial role for indoor scene understanding, especially in robotics applications. We propose a new domain invariant Real-to-Simulation (Real-Sim) fine-tuning strategy for SAM. We use object images and ground truth data collected from Ai2Thor simulator during fine-tuning (real-to-sim). To allow our Segment Any Object Model (SAOM) to work in the “everything” mode, we propose the novel nearest neighbour assignment method, updating point embeddings for each ground-truth mask. SAOM is evaluated on our own dataset collected from Ai2Thor simulator. SAOM significantly improves on SAM, with a $28 \%$ increase in mIoU and a $25 \%$ increase in mAcc for 54 frequently-seen indoor object classes. Moreover, our Real-to-Simulation fine-tuning strategy demonstrates promising generalization performance in real environments without being trained on the real-world data (sim-to-real). The dataset and the code are available here.

Embodied navigation requires robots to understand and interact with the environment based on given tasks. Vision-Language Navigation (VLN) is an embodied navigation task, where a robot navigates within a previously seen and unseen environment, based on linguistic instruction and visual inputs. VLN agents need access to both local and global action spaces; former for immediate decision making and the latter for recovering from navigational mistakes. Prior VLN agents rely only on instruction-viewpoint alignment for local and global decision making and back-track to a previously visited viewpoint, if the instruction and its current viewpoint mismatches. These methods are prone to mistakes, due to the complexity of the instruction and partial observability of the environment. We posit that, back-tracking is sub-optimal and agent that is aware of its mistakes can recover efficiently. For optimal recovery, exploration should be extended to unexplored viewpoints (or frontiers). The optimal frontier is a recently observed but unexplored viewpoint that aligns with the instruction and is novel. We introduce a memory-based and mistake-aware path planning strategy for VLN agents, called StratXplore, that presents global and local action planning to select the optimal frontier for path correction. The proposed method collects all past actions and viewpoint features during navigation and then selects the optimal frontier suitable for recovery. Experimental results show this simple yet effective strategy improves the success rate on two VLN datasets with different task complexities.

Understanding of scene changes is crucial for embodied AI applications, such as visual room rearrangement, where the agent must revert changes by restoring the objects to their original locations or states. Visual changes between two scenes, pre- and post-rearrangement, encompass two tasks: scene change detection (locating changes) and image difference captioning (describing changes). While previous methods, focused on sequential 2D images, have addressed these tasks separately, it is essential to emphasize the significance of their combination. Therefore, we propose a new Scene Change Understanding (SCU) task for simultaneous change detection and description. Moreover, we go beyond change language description generation and aim to generate rearrangement instructions for the robotic agent to revert changes. To solve this task, we propose a novel method - EmbSCU, which allows to compare instance-level change object masks (for 53 frequently-seen indoor object classes) before and after changes and generate rearrangement language instructions for the agent. EmbSCU is built on our Segment Any Object Model (SAOMv2) - a fine-tuned version of Segment Anything Model (SAM), adapted to obtain instance-level object masks for both foreground and background objects in indoor embodied environments. EmbSCU is evaluated on our own dataset of sequential 2D image pairs before and after changes, collected from the Ai2Thor simulator. The proposed framework achieves promising results in both change detection and change description. Moreover, EmbSCU demonstrates positive generalization results on real-world scenes without using any real-life data during training. The dataset and the code are available here.

In the dynamic landscape of cyber threats, multi-stage malware botnets have surfaced as significant threats of concern. These sophisticated threats can exploit Internet of Things (IoT) devices to undertake an array of cyberattacks, ranging from basic infections to complex operations such as phishing, cryptojacking, and distributed denial of service (DDoS) attacks. Existing machine learning solutions are often constrained by their limited generalizability across various datasets and their inability to adapt to the mutable patterns of malware attacks in real world environments, a challenge known as model drift. This limitation highlights the pressing need for adaptive Intrusion Detection Systems (IDS), capable of adjusting to evolving threat patterns and new or unseen attacks. This paper introduces MalBoT-DRL, a robust malware botnet detector using deep reinforcement learning. Designed to detect botnets throughout their entire lifecycle, MalBoT-DRL has better generalizability and offers a resilient solution to model drift. This model integrates damped incremental statistics with an attention rewards mechanism, a combination that has not been extensively explored in literature. This integration enables MalBoT-DRL to dynamically adapt to the ever-changing malware patterns within IoT environments. The performance of MalBoT-DRL has been validated via trace-driven experiments using two representative datasets, MedBIoT and N-BaIoT, resulting in exceptional average detection rates of 99.80% and 99.40% in the early and late detection phases, respectively. To the best of our knowledge, this work introduces one of the first studies to investigate the efficacy of reinforcement learning in enhancing the generalizability of IDS.