Semantic communication is a neural network-based paradigm for future 6G networks that enhances transmission efficiency by conveying semantic features instead of raw data. Task-oriented semantic communication further optimizes resource usage by focusing solely on task-relevant features without sacrificing accuracy. These systems, however, are vulnerable to model inversion attacks, in which adversaries use crafted inputs to probe the transmitter, gather responses, and train inverse models that can reconstruct original inputs from the transmitter’s output, leading to privacy risks. Traditional defenses typically focus on minimizing visual similarities between the original and reconstructed data, but this approach is inadequate in semantic communication scenarios, as visual differences do not necessarily ensure semantic divergence. In this paper, we propose a novel diffusion-based semantic communication framework that effectively balances semantic compression with robustness against channel noise, ensuring stable performance across diverse channel conditions. Additionally, we introduce a new evaluation metric specifically designed to quantify the extent of semantic information leakage, thereby providing a more accurate measure of adversarial threats in task-oriented scenarios. Experimental results on the MNIST dataset demonstrate that our proposed framework achieves a classification accuracy improvement of 25.4% over adversarial reconstructions. Our findings also reveal a critical disconnect between traditional image-quality metrics and the actual leakage of task-relevant semantic information.

In SplitFed learning (SFL), a global model is split into two segments, where distributed clients train the first segment in a federated manner and a main server trains the other. Existing studies focus on algorithm development but ignore the important issue of incentives, without which self-interested clients may be unwilling to participate. We fill this gap by presenting a first incentive study in SFL. One challenge is that the design requires an understanding of how clients' participation affects the model performance. To this end, we provide a first convergence analysis for SFL considering partial client participation to guide the mechanism design. Another challenge is that monetary payment may not be viable for large distributed systems. To this end, we propose a model-versioning mechanism where the main server assigns different versions of models (of different qualities) to clients as incentives. The design is further complicated by clients' multi-dimensional private information. To this end, we design the model-versioning mechanism so that it decouples clients' decisions and admits a weakly dominant strategy at equilibrium. We prove that our mechanism is feasible, effective, and incentive compatible. Experimental results show that our mechanism greatly improves client participation and model accuracy compared to a benchmark.

Knowledge distillation (KD) can enable collaborative learning among distributed clients that have different model architectures and do not share their local data and model parameters with others. Each client updates its local model using the average model output/feature of all client models as the target, known as federated KD. However, existing federated KD methods often do not perform well when clients’ local models are trained with heterogeneous local datasets. In this paper, we propose Federated knowledge distillation enabled by Adversarial Learning (FedAL) to address the data heterogeneity among clients. First, to alleviate the local model output divergence across clients caused by data heterogeneity, the server acts as a discriminator to guide clients’ local model training to achieve consensus model outputs among clients through a min-max game between clients and the discriminator. Moreover, catastrophic forgetting may happen during the clients’ local training and global knowledge transfer due to clients’ heterogeneous local data. Towards this challenge, we design the less-forgetting regularization for both local training and global knowledge transfer to guarantee clients’ ability to transfer/learn knowledge to/from others. Experimental results show thatFedALand its variants achieve higher accuracy than other federated KD baselines.

Routing is critical for maximizing the performance of Dragonfly networks. It decides how the packets are forwarded from their source nodes to their destination nodes. Considering the traffic pattern varies over time in real networks, adaptive routing is desirable. Existing adaptive routing algorithms employ local information to make dynamic routing decisions, which have shown significant limitations since the local information typically fails to reflect the global network condition. In this paper, inspired by the hierarchical topology of Dragonfly, we develop a hierarchical Reinforcement Learning (RL) algorithm named Q-hierarchical for Dragonfly networks. Q-hierarchical learns to adapt from data without the need to model the traffic pattern. It simplifies the complexity of traditional RL-based routing by routing in a hierarchical manner, i.e., inter-group routing and intra-group routing. Hence it can be applied on a large network. We also develop a fast and effective learning strategy for the hierarchical RL. The performance of Q-hierarchical is evaluated through comprehensive tests on two Dragonfly topologies. The results show that our approach provides comparable performance under uniform random traffic pattern and outperforms some routing algorithms in terms of averaging packet delay and affordable load under adversarial traffic pattern.