Alternative Multi-Phase Training with Mask Attack for Digital Semantic Communications
Mingze Gong, Shuoyao Wang, Suzhi Bi · 2024
Semantic communication (SemComm) has emerged as new paradigm shifts. However, most existing SemComm systems transmit continuously distributed signals in analog fashion, being incompatible with the current digital communication framework. In this paper, we propose a digital SemComm system for image transmission, named AMP-SC. To address the non-differentiability issues, we propose an alternative optimization approach for the encoder and decoder, called the alternative multi-phase training strategy (AMP). AMP contains three training phases: feature extraction (FE), robustness enhancement (RE), and training-testing alignment (TTA), respectively. In particular, the alternation between solely updating the decoder and jointly optimizing codecs in RE, enables parameter updating in both the encoder and decoder, despite the non-differentiable digital communication processes in between. To further boost robustness, we investigate a mask-attack (MATK) in RE to simulate significant bit-flipping effect in a differentiable manner. From the neural network perspective, we explore an information restoration network called IRSNet to restore received analog features with special correlations, mitigating the impact of information loss caused by bit-flipping. To address the training-testing inconsistency introduced by MATK, we employ an additional TTA phase to fine-tune the decoder without MATK. Comparing with the representative benchmark, AMP-SC achieves a 1.24 dB higher average performance across various signal-to-noise ratios.