Neural Demodulation for Anti-Eavesdropping Embedded Waveforms in IoT Networks

Internet of Things (IoT) devices often deliver sensitive sensing and control data over wireless links that are inherently broadcast, making passive over-the-air eavesdropping a persistent threat even for simple point-to-point transmissions. Many physical-layer security (PLS) techniques, however, rely on accurate channel knowledge, additional spatial degrees of freedom, or strong secrecy assumptions that are often hard to guarantee in practical IoT links. In this article, we propose an anti-eavesdropping secure transmission scheme by deliberately embedding controllable artificial interference into a conventional binary phase shift keying (BPSK) waveform on the same carrier, thereby forming an interference-embedded waveform transmitted over an additive white Gaussian noise (AWGN) channel. To exploit the common architectural asymmetry in IoT systems, the legitimate receiver located at an IoT gateway/edge node employs an anti-eavesdropping waveform demodulator (AEWD), a ResNet-self-attention-bidirectional long short-term memory (BiLSTM) neural demodulator trained end-to-end to recover bits directly from raw in-phase/quadrature (I/Q) samples without explicit interference parameter estimation. Under identical channel conditions, conventional model-based receivers that perform deterministic interference suppression followed by demodulation incur substantial bit error rate (BER) degradation and frequently exhibit interference-limited error floors. Extensive simulations across a wide range of signal-to-noise ratio (SNR) and signal-to-interference ratio (SIR) demonstrate that AEWD consistently outperforms classical baselines under both single-tone and multitone nonstationary interference. The results suggest that waveform-level interference embedding combined with a dedicated neural demodulator can create a practical receiver performance gap to strengthen confidentiality for IoT communications.