ENFCL: Self-Supervised Contrastive Learning Framework of Electric Network Frequency Representations for Generalizable Audio Tampering Detection

Digital audio tampering detection is essential for preserving information authenticity, sustaining societal trust, and safeguarding public safety. However, most electric network frequency (ENF)–based detectors are trained in a fully supervised manner and therefore depend on large-scale labeled datasets, which are costly to obtain and often unavailable in real-world deployments. This heavy reliance on annotations limits representation discriminability and weakens generalization across unseen recording conditions and tampering patterns. To address these challenges, we propose ENF-based contrastive learning (ENFCL), a self-supervised framework that introduces contrastive pretraining for ENF-based audio tampering detection. ENFCL learns discriminative ENF representations from unlabeled audio by constructing diverse positive and negative sample pairs, thereby improving robustness and generalization without manual annotations. For pretraining, we design a multiscale encoder, multiscale ENF feature encoder with transformer (MSENFformer) to capture tampering cues at different resolutions. MSENFformer integrates three complementary subnetworks, zero-order phase deep residual network (ZPDR-Net), first-order phase deep representation network (FPDR-Net), and multiple-harmonics deep aggregation network (MHDA-Net) to model ENF characteristics across phase orders and harmonic components. A transformer module further aggregates global contextual information to enrich hierarchical feature representations. To mitigate the adaptation gap during fine-tuning, we introduce grouped convolution and dual-attention–based selective kernel network (GCDA-SKNet), which adaptively fuses multiscale features via dynamic selective kernels, enhancing discrimination for downstream classification. Experiments demonstrate that ENFCL achieves 94.00% accuracy on ENF-EDIT1, substantially outperforming existing state-of-the-art methods. Across ENF-EDIT1, ENF-EDIT2, and ENF-HBUT, ENFCL remains consistently strong under diverse data augmentation settings, confirming its robustness and generalization in practical audio tampering detection scenarios.