Fisher-Robust Information Bottleneck for Task-Oriented Communication with Noisy Data

In this paper, we consider task-oriented commu-nication for edge AI inference, where the edge device extracts features from input data with potential noise and then transmits the encoded representations to the edge server over the physical channel for inference tasks. However, it remains an open problem to balance the tradeoff between inference accuracy and robustness in the presence of both data noise and channel noise. To address this issue, we propose a robust encoding framework for task-oriented communication, named Dual-Robust Encoding (DRE), where a robust encoding framework is developed, named Fisher-Robust Information Bottleneck (FR-IB), to balance the tradeoff between relevance and robustness. We derive a tractable variational upper bound of the FR-IB objective function using the variational approximation to overcome the computational intractability of mutual information. Experiments on image classification tasks demonstrate that DRE outperforms baseline methods, showing improved robustness to data and channel variations.