Adversarial Fine-Tuning on Speech Foundation Model with Vulnerable Attention Consistency Regularization for Robust Speech Recognition
📄 Adversarial Fine-Tuning on Speech Foundation Model with Vulnerable Attention Consistency Regularization for Robust Speech Recognition #语音识别 #语音大模型 #预训练 #对抗样本 #鲁棒性 ✅ 7.5/10 | 前25% | #语音识别 | #对抗样本 | #语音大模型 #预训练 学术质量 6.5/7 | 选题价值 1.5/2 | 复现加成 0.5 | 置信度 高 👥 作者与机构 第一作者:Yanyun Wang (The Hong Kong University of Science and Technology (Guangzhou)) 通讯作者:Li Liu (The Hong Kong University of Science and Technology (Guangzhou), avrillliu@hkust-gz.edu.cn) 作者列表:Yanyun Wang (The Hong Kong University of Science and Technology (Guangzhou)), Baoyuan Wu (The Chinese University of Hong Kong, Shenzhen; Shenzhen Loop Area Institute), Li Liu (The Hong Kong University of Science and Technology (Guangzhou)) 💡 毒舌点评 亮点:这篇工作敏锐地抓住了“防御SFM时,不能像对待传统模型那样容忍精度大幅下降”这一核心矛盾,通过CKA分析定位脆弱层并设计了针对性的双重正则化(注意力散度和特征相似性),思路清晰且可解释性强。短板:实验基本局限于Whisper模型在LibriSpeech一个数据集上的表现,对于SFM在多语言、多噪声环境下的泛化能力验证不足,使得“SOTA”的宣称在更大范围内略显底气不足。 ...