Deep learning has advanced in fields like computer vision and natural language processing but still faces reliability challenges, including adversarial attacks that mislead models and threaten security-sensitive applications. Generative models help counter such threats by purifying adversarial images. Among them, energy-based models (EBMs) are effective for adversarial purification. Our approach utilizes EBMs for purification of inputs into a classifier while incorporating adversarial examples of that classifier into training procedure of the EBM. The proposed EBM training technique outperforms vanilla training procedure by 10.52% and 20.48% on MNIST and FashionMNIST, respectively, for the purification task. In our experiment setting, robust accuracy, defined as the accuracy of adversarial examples in the test set after the purification process, achieves 48% against benchmark AutoAttack (L-infinity, epsilon = 8/255) on CIFAR-10, improving the state-of-the-art purification method using EBM by 1.1%, with a clean accuracy of 70%.
In secure multi-party computation, multiple data sources are involved, and a master needs to compute a function over them. This computation is often offloaded to processing nodes, some of which may collude in an attempt to extract private information. Therefore, ensuring the privacy of the data is crucial. In this work, we assume that the data are large matrices, which need to be partitioned into smaller pieces for storage in the processors. Additionally, there is a limited-capacity communication link between the data sources and processors. Consequently, we aim to upload all the data in a single round, while allowing processors to communicate with one another. We propose an approach to matrix multiplication that incorporates Strassen algorithm, enabling us to compute the function with fewer processors while preserving privacy against both the processing nodes and the master. Moreover, our scheme efficiently handles a distributed batch matrix multiplication setting, where an identical function needs to be applied to different datasets.
In multi-path networks, there is the possibility of out-of-order packet receptions at the destination, i.e., a packet generated sooner than another one can be received later, which is the main challenge in analyzing the age-related metrics in multi-path networks. Most of the previous works on multi-path networks have only focused on studying the average age of information (AoI). However, analyzing the average peak AoI (PAoI), another age-related metric, can be even more challenging since we need to know when a local peak occurs in the AoI function. In this work, for a general multi-path network, we find the relation of the average PAoI based on some characteristics of the ages of the incoming links at the destination. As a case study, we derive the average PAoI in a network of two parallel preemptive queues. In the symmetric case, where sampling and transmission rates are identical for both queues, we find that based on the amounts of sampling and transmission rates, the average PAoI is between 38.33 to 50 percent smaller compared to the scenario with only one queue.
Critical massive machine-type communications (cmMTC) in 5G networks face significant challenges due to congestion and resource contention. This paper proposes a random back-off mechanism within an adaptive back-off window and an adaptive retransmission strategy designed to enhance resource allocation in cmMTC scenarios through a groupingbased approach. Unlike traditional random-access mechanisms, our method mitigates access failures by incorporating a random back-off within an adaptive back-off window and an adaptive retransmission mechanism for devices that initially fail to access the network. By optimizing the back-off and retransmission process, the strategy significantly reduces access latency, enhances reliability, and minimizes collisions. Through analytical modeling and extensive simulations, we demonstrate that our approach outperforms conventional random-access methods in terms of collision probability, access success rate, and average transmission delay. The results confirm that our load adaptive selection strategy effectively balances network load while maintaining high connection rates, making it a robust and scalable solution for future 5G-enabled IoT applications.
This paper investigates the joint optimization of the base station precoding matrix and the beyond-diagonal reconfigurable intelligent surface (BD-RIS) scattering matrix in an integrated sensing and communication (ISAC) system that employs rate-splitting multiple access (RSMA). To tackle the inherent nonconvexity and NP-hardness of the optimization problem and channel estimation uncertainties, we propose a novel deep reinforcement learning (DRL) framework, termed SAC-QA. The proposed SAC-QA algorithm leverages action quantization to discretize the continuous high-dimensional action space and incorporates a low-rank approximation with sparse correction for efficient BD-RIS matrix generation. Simulation results demonstrate that the proposed approach significantly improves the worst-case performance compared to conventional algorithms, thereby highlighting its potential for enabling robust and adaptive ISAC systems in next-generation wireless networks.
