Addressing the Diglossic Gap in Arabic audio deepfake detection with a multi-dialect benchmark and multi-representation fusion detector

Abstract Arabic audio deepfake detection is challenged by the interaction of within-language dialectal variation, prosody, recording conditions, codec history, and rapidly evolving synthesis technologies. We characterize this compound target-domain shift through the Diglossic Gap , a structured taxonomy spanning dialectal/phonetic, prosodic, channel/acquisition, and synthesis-source variation. We introduce DeepShield, a 5030-file Arabic audio corpus comprising 2375 genuine and 2655 synthetic recordings spanning voice cloning, text-to-speech, and voice conversion. The experimental protocol uses a balanced 1600-file target-adaptation partition and a locked, file-disjoint 2089-file independent test partition, with an additional 1341 recordings reserved outside the active protocol. DeepShield Fusion integrates SincNet/ECAPA-TDNN, wav2vec 2.0, and EfficientNetV2-S + BiGRU evidence through validation-locked score fusion and achieves 91.00% accuracy, 9.20% equal error rate (EER), and 0.9652 area under the receiver operating characteristic curve (AUC) using the opening 4 s of each recording. A duration-stratified ablation further verifies the contribution of padding-aware Smart Attention-Mask Pooling, reducing wav2vec 2.0 EER from 17.90% to 17.25%. Comprehensive comparative evaluation against RawNet2, WavLM-Base+, and LFCC-LCNN yields EERs of 8.31%, 6.49%, and 5.19%, respectively. Building on the complementary behavior of self-supervised and spectral representations, DeepShield-Fusion+ combines a partially fine-tuned WavLM-Base+ branch with LFCC-LCNN through development-trained logistic stacking and achieves 97.42% accuracy, 2.69% EER, and 0.9964 AUC. Relative to LFCC-LCNN, Fusion+ reduces EER by 2.50 percentage points (95% CI [ $$-3.56,-1.75$$ ]). Robustness analyses additionally quantify performance under controlled additive-noise, bandwidth-limiting, and standardized-transcoding conditions. Under uniform mono 16-kHz Opus transcoding at 24 kb/s, Fusion+ retains 95.64% accuracy, 3.36% EER, and 0.9948 AUC. Collectively, the results establish DeepShield as a rigorous Arabic target-domain benchmark and demonstrate the effectiveness of padding-aware self-supervised modeling and heterogeneous learned fusion for localized audio deepfake detection.