F.2.7. Optimization
alireza Omidi nasab; Sajad Bastami; Rojiar Pir Mohammadiani; Mohammad Bagher Dowlatshahi; Seyedeh Zahra Mousavi
Abstract
Deep Neural Networks (DNNs) are increasingly deployed in safety-critical domains such as autonomous driving, healthcare, finance, and natural language processing, yet they remain vulnerable to adversarial attacks—subtle manipulations that can cause confident misclassifications or misleading predictions. ...
Read More
Deep Neural Networks (DNNs) are increasingly deployed in safety-critical domains such as autonomous driving, healthcare, finance, and natural language processing, yet they remain vulnerable to adversarial attacks—subtle manipulations that can cause confident misclassifications or misleading predictions. This fragility poses a major barrier to building secure and trustworthy AI systems. Conventional defenses, including adversarial training and heuristic detection, often struggle to balance robustness, adaptability, and computational cost. To overcome these limitations, we propose a hybrid adaptive defense framework that unifies Ant Colony Optimization (ACO) with Reinforcement Learning (RL). ACO efficiently explores the high-dimensional space of defense hyperparameters to find globally optimal configurations, while RL enables dynamic, context-aware adaptation of defense strategies in real time. The proposed ACO-RL framework was rigorously evaluated across six diverse benchmark datasets spanning multiple data modalities: MNIST and CIFAR-10 (vision), IMDB and AG News (text), and Cora and Reddit-Binary (graph). Experimental results show that ACO-RL consistently enhances robustness against a wide spectrum of adversarial attacks, outperforming several state-of-the-art baselines. These findings highlight a promising pathway toward developing resilient, cross-domain AI systems capable of defending against evolving adversarial threats.
B.3. Communication/Networking and Information Technology
Sajad Bastami; Mohammad Bagher Dowlatshahi; Rojiar Pir Mohammadiani; Seyedeh Zahra Mousavi
Abstract
Multi-agent reinforcement learning (MARL) is a key paradigm for coordination in robotics, autonomous systems, and distributed control. However, existing MARL methods face fundamental limitations in scalability, adaptability to dynamic environments, and stability under evolving interactions. To address ...
Read More
Multi-agent reinforcement learning (MARL) is a key paradigm for coordination in robotics, autonomous systems, and distributed control. However, existing MARL methods face fundamental limitations in scalability, adaptability to dynamic environments, and stability under evolving interactions. To address these challenges, we propose Adaptive Graph-Transformer Reinforcement Learning (AGTRL), a framework integrating graph-based relational modelling with transformer attention for adaptive coordination in large-scale multi-agent systems. AGTRL unifies graph-based perception and attention-based coordination in an end-to-end pipeline, encoding role information and adaptively weighting interactions by context. This combination, missing in prior MARL methods, bridges scalability and robustness in dynamic environments. AGTRL constructs a dynamic graph of agent relationships and uses multi-head self-attention to prioritize relevant interactions in real-time, ensuring robust performance under perturbations. We evaluate robustness under communication dropout (up to 40% link removal) and dynamic edge removal, measuring performance via episode reward and win rate. The framework incorporates an adaptive stability-performance trade-off mechanism that maintains learning efficacy in the presence of communication constraints and environmental uncertainty. We introduce a graph-enhanced policy architecture that jointly optimizes individual agent policies and inter-agent coordination through attention-weighted message passing. Comprehensive evaluations on benchmark environments—including StarCraft II micromanagement scenarios, cooperative navigation (Spread), and adversarial tasks (Predator-Prey)—demonstrate that AGTRL achieves superior sample efficiency, scalability, and robustness compared to state-of-the-art MARL baselines. Experimental results show AGTRL improves convergence speed by 32% on average and maintains stable performance with up to 40% communication dropout, establishing its viability for real-world deployment in dynamic multi-agent domains.