Similarity Choice and Negative Scaling in Supervised Contrastive Learning for Deepfake Audio Detection

Researchers Jaskirat Sudan, Hashim Ali, Surya Subramani, and Hafiz Malik from the University of Michigan-Dearborn published a controlled study on supervised contrastive learning (SupCon) for audio deepfake detection, a domain where SupCon's specific impact was previously underexplored. Using the wav2vec2 XLS-R (300M) model, they systematically varied two key SupCon components: similarity measure (cosine vs angular similarity derived from the hyperspherical angle) and negative scaling via a warm-started global cross-batch queue. The two-stage pipeline first fine-tuned the encoder and projection head with SupCon, then froze them to train a linear classifier with binary cross-entropy loss.

Trained on the ASVspoof 2019 LA dataset and evaluated across ASV19 eval, ITW, and ASVspoof 2021 DF/LA benchmarks, Cosine SupCon with a delayed queue achieved the best results: ITW equal error rate (EER) of 8.29% and pooled EER of 4.44%. Interestingly, angular similarity performed strongly without queued negatives (ITW 8.70%), suggesting reduced reliance on large negative sets—a finding that could simplify training pipelines. This work provides clear guidance for practitioners building robust deepfake audio detectors, emphasizing the importance of similarity choice and negative sampling strategy in contrastive learning.