M.S. in Artificial Intelligence ('26) · University of Michigan
Interested in what signals encode about the physical world and how to build models that learn that structure robustly.
MS in AI from the University of Michigan, researching audio deepfake detection at the ISSF Lab under Prof. Hafiz Malik. My work focuses on the latent representations of speech produced by SSL models: how similarity function choice and negative scaling shape the geometry of those representations, and what that geometry reveals about the boundary between real and synthetic voice. I use supervised contrastive learning to reshape the embedding space of large speech encoders (XLS-R, WavLM, HuBERT) for deepfake detection, and I am interested in what different layers encode about the physical structure of speech.
My background spans representation learning, contrastive and self-supervised methods, and signal processing. On the computer vision side, I have worked on semantic segmentation, behavioral cloning agents, and Siamese networks.
Research
A controlled study of what actually drives out-of-distribution generalization in SSL-based deepfake detection. Over a frozen XLS-R-300M, a contrastive projection head (Stage 1) feeds a linear classifier (Stage 2); I sweep similarity geometry (cosine vs. geodesic), negative-queue size, and temperature to isolate each factor's effect. 0.25% EER on ASVspoof 2019, 8.3% on In-the-Wild.
A from-scratch rebuild of a two-stage SSL deepfake detector (Wav2Vec2 / WavLM): Stage 1 compresses style and linguistic structure, Stage 2 classifies with waveform augmentation. Adds per-speaker embedding diagnostics and UMAP/t-SNE views that expose entity-level failures hidden by aggregate EER.
The first passive visible-light tag for door access: layered transparent tapes exploit polarized birefringence to produce 3D, position-dependent color patterns that can't be cloned. A real-time CV pipeline (SIFT → FLANN → RANSAC, no deep learning) hits 90.5% accuracy at 0.5m, <100ms latency, for a $0.20 tag.
Projects
A deployed web app that flags AI-generated speech using a spectrogram autoencoder: audio that reconstructs poorly is scored as synthetic. Accepts file uploads, batches, and YouTube links, renders the spectrogram, and produces a plain-language AI summary of each verdict.
An autoencoder trained only on real speech flags deepfakes by reconstruction error, synthetic audio reconstructs poorly. Compares CNN vs. ViT-Tiny backbones and two spectrogram masks. Best config: 15.5% EER on ASVspoof 2019, 13.1% on In-the-Wild, fully zero-shot.
End-to-end behavioral cloning: a CNN-LSTM learns to drive from 50K frame-action pairs, combining spatial features with temporal memory. Runs closed-loop at ≤60ms latency, hitting 94% action accuracy and completing full laps autonomously.
Tackles the day-to-night domain gap on BDD100K by pretraining on daytime then fine-tuning on night. Lifting an Xception U-Net from 0.08 to 0.90 Dice. Benchmarked against MobileNetV2 U-Net to map the accuracy/compute tradeoff.
A local Tkinter interface for Meta's SAM (ViT-H/L/B) with point, box, and text prompts and real-time mask preview, full interactive segmentation on consumer hardware, no cloud.
An autoencoder you can reshape live: encoder/decoder depth and latent dimension are adjustable at runtime, with a GUI showing reconstructions update as you change the bottleneck, a hands-on tool for building intuition about compression.
Estimates latitude/longitude from a photo of the night sky by matching it against a star catalog. Combines template matching, SIFT/ORB homography, and a grid search for robustness to rotation and scale.
Experience
Research Assistant — Speech ML / Audio Anti-Spoofing
Advisor: Prof. Hafiz Malik
Research Assistant — Computer Vision & ML Systems
Advisor: Prof. Xiao Zhang
Research Assistant — Signal Processing + ML
Advisor: Prof. Narendranath Patra
Publications
Jaskirat Sudan, Hashim Ali, Surya Subramani, Hafiz Malik
A controlled study of how similarity-function choice and negative scaling drive out-of-distribution generalization in SSL-based audio deepfake detection. A two-stage pipeline over a frozen Wav2Vec2-XLS-R-300M backbone, supervised contrastive projection head, then linear classifier. Achieves 0.25% EER on ASVspoof 2019 and 8.3% EER on In-the-Wild under single-dataset training.
Jaskirat Sudan, Fatima Qasem, Hasky E Fynn, Fatima Mohammed, Ashwin Sarvadey, Tian Xie, Ang Li, Xiao Zhang
Low-cost, privacy-preserving door access control via visible-light backscatter. Polarized birefringence tags create 3D, position-dependent color patterns as unclonable optical keys. The <$0.20 prototype achieves 90.5% authentication accuracy at 0.5m while eliminating cloning, replay, and privacy attack surfaces.
Hasky Fynn, Jaskirat Sudan, Fatima Qasem, Fatima Mohammed, Xiao Zhang
Companion demo paper presenting the live ViKey system, the first visible light backscatter-based door access control system using polarized birefringence to generate 3D position-dependent color patterns as keys, enabling robust and contactless authentication.
Safdar Sardar Khan*, Jaskirat Singh Sudan*, Anuj Pathak*, Rakesh Pandit, Pinky Rane, Ashish Kumar Kumawat (* equal contribution)
Systematic survey of 25+ EEG denoising methods across statistical, ICA-based, wavelet-based, and DL-based approaches for BCI applications. Hybrid ICA + wavelet pipelines consistently outperform single-method approaches for preserving neural signal quality in real-time pipelines.
Blog
How radio telescopes see what eyes can't, revealing pulsars, black holes, and interstellar gas to map the invisible universe.
How an early-warning algorithm and crucial human judgment prevented a Cold War false alarm from escalating into catastrophe.