Enhancing Efficiency and Performance in Deepfake Audio Detection through Neuron-level Dropin & Neuroplasticity Mechanisms

Researchers developed novel 'dropin' and 'plasticity' algorithms inspired by brain neuroplasticity to improve deepfake audio detection efficiency. The methods dynamically adjust neuron counts in model layers, achieving up to 66% reduction in error rates while improving computational efficiency across multiple architectures including ResNet and Wav2Vec.

• →New algorithms inspired by mammalian brain neuroplasticity can dynamically adjust neuron counts to optimize model performance without full retraining.
• →The dropin approach achieved up to 39% relative reduction in Equal Error Rate while improving computational efficiency.
• →The plasticity method demonstrated up to 66% relative reduction in Equal Error Rate on deepfake detection datasets.
• →Unlike existing methods limited to attention-based architectures, these algorithms work across diverse models including ResNet and Wav2Vec.
• →The approach addresses the computational bottleneck of simply scaling model parameters through additional layers.