My PhD thesis built defenses for speech systems under worst-case threat models (full attacker knowledge, zero defender knowledge), as the lead of JHU’s top-ranked blue team across the DARPA GARD and RED programs.

• Adversarial ASR: tandem adversarial fine-tuning + denoiser defense cut WER degradation 45% against PGD-500 attacks (Joshi et al., 2022) (Interspeech 2022, Oral).
• Data poisoning: an unsupervised KMeans defense over DINO embeddings dropped attack success from 99% → 0.25% (Thebaud et al., 2023) (ASRU 2023).
• Speaker recognition: ParallelWaveGAN vocoder preprocessing gave ~41% average absolute robustness gain (Joshi et al., 2021) (IEEE TIFS, IF 7.2).
• Attack forensics: ~90% accuracy classifying known attacks and victim models (Joshi et al., 2024) (Odyssey 2024, Oral).

Dissertation: Securing Speech Systems Against Adversarial and Poisoning Attacks (JHU, 2025).