Jaskirat Singh Sudan

M.S. in Artificial Intelligence ('26) · University of Michigan–Dearborn

Representation Learning Audio Deepfake Detection Speech + Vision ML

GitHub · LinkedIn · Google Scholar · YouTube

MS AI student at the University of Michigan–Dearborn (graduating May 2026), researching contrastive representation learning for audio deepfake detection at the ISSF Lab under Prof. Hafiz Malik. My thesis is a controlled study of how similarity function choice and negative scaling affect OOD generalization in SSL speech models for deepfake detection tasks achieving 0.25% EER on ASVspoof 2019 and 8.3% on In-the-Wild. My background spans computer vision (segmentation, behavioral cloning, Siamese networks), signal processing (GMRT radio data, star navigation), and a published IEEE paper on optical authentication (ViKey, IEEE MASS 2025). I'm interested in what signals encode about the physical world and how to build models that learn that structure robustly.

Open to full-time roles in audio ML, speech AI, and representation learning starting May 2026.

Research

Thesis — Contrastive Audio Deepfake Detection

Master's Thesis Aug 2025–Present Representation Learning

Supervised Contrastive Learning with Cross-Batch Negatives for Deepfake Audio Detection

Controlled study of cosine vs. geodesic similarity and cross-batch negative scaling (|Q|=2048) over frozen XLS-R-300M. Stage 1: SupCon projection head. Stage 2: linear BCE classifier. Ablations: temperature sweep, queue-size sweep, UMAP geometry. Evaluated across FakeXpose and FamousFigures OOD corpora.

0.25% EER · ASVspoof 2019 8.3% EER · In-the-Wild 0.8% / 6.6% · pooled corpus

Presentation → Poster → Code →

SLIM — two-stage SSL deepfake detection replication

Oct 2025 Self-Supervised Learning

Speaker-Specific SLIM Replication (From-Scratch)

From-scratch replication of a two-stage SSL deepfake detection framework using Wav2Vec2 and WavLM backbone variants. Stage 1 learns compressed style/linguistic representations; Stage 2 trains a classifier with waveform augmentations (noise, reverb, SpecAugment). Built per-speaker embedding diagnostics and UMAP/t-SNE visualizations to surface entity-level failure modes invisible in aggregate EER.

Code: Private (ISSF Lab)

ViKey — visible-light backscatter authentication

IEEE MASS 2025 Oct 2025 Co-Primary Author Published

ViKey: Secure Door Access Control Using Passive Visible Light Tags

Proposed the first passive visible light tag for door access control, exploiting polarized birefringence in layered transparent tapes to generate 3D position-dependent color patterns as unclonable keys. Conducted a systematic case study of tag shape, tape layers, and stacking sequences, and designed a real-time CV pipeline using per-channel SIFT, FLANN matching, RANSAC geometric verification, and temporal consistency checks achieving 90.5% accuracy at 0.5m and <100ms latency without any deep learning.

90.5% accuracy @ 0.5m <100ms latency $0.20 tag cost

Paper → Code →

Projects

Apr 2025 · Representation Learning

Contrastive Learning with Siamese Network

128-D MNIST embedding space with 99% pair accuracy; Tkinter + Plotly GUI for few-shot classification and 3D PCA visualization.

Code → Demo →

Apr 2025 · Behavioral Cloning

Self-Driving Mario Kart (CNN-LSTM)

CNN-LSTM trained on 50K frame-action pairs at ≤60ms emulator latency. 94% action accuracy, autonomous lap completion.

Code → Slides →

Nov 2024 · Transfer Learning

Low-Light Semantic Segmentation

Xception U-Net Dice 0.08→0.90 on BDD100K nighttime. MobileNetV2 U-Net benchmarked for accuracy/compute tradeoff.

Code → Slides →

Dec 2024 · Diffusion / Multimodal

Speech→Image with Latent Diffusion

Whisper ASR + DreamBooth-fine-tuned Stable Diffusion v2 with mixed-precision training and prior preservation loss.

Code → Talk → Demo →

Feb 2025 · Image Segmentation

Segment Anything Desktop GUI

Tkinter interface for SAM (ViT-H/L/B). Point, box, and text prompts with real-time mask visualization on consumer hardware.

Code →

Feb 2024 · Representation Learning

Scalable Convolutional Autoencoder (SCA)

Configurable encoder/decoder depth with live reconstruction GUI. Adjustable latent dimension at runtime.

Code → Demo →

2023 · Computer Vision + Astronomy

Star-Based Navigation

Lat/lon estimation via star-catalog matching. Template, SIFT/ORB homography, and grid-based strategies combined for rotation/scale robustness.

Code →

2023 · Computer Vision + HCI

Air-Draw (Gesture-Controlled Whiteboard)

MediaPipe + OpenCV virtual whiteboard for live hand-gesture annotation in video calls without a keyboard or mouse.

Code → Demo →

Experience

ISSF Lab, University of Michigan–Dearborn

Aug 2025 – Present · Dearborn, MI

Research Assistant — Speech ML / Audio Anti-Spoofing

Advisor: Prof. Hafiz Malik

–Developing MS thesis on contrastive representation learning for audio deepfake detection: controlled study of cosine vs. geodesic similarity and cross-batch negative scaling over frozen XLS-R-300M. Best results: 0.25% EER (ASVspoof 2019 eval), 8.3% EER (In-the-Wild); 0.8% / 6.6% with pooled corpus training across ASVspoof 2019, ASVspoof 5, and MLAAD.
–Replicated SLIM (two-stage SSL deepfake detection pipeline) from scratch using Wav2Vec2/WavLM; built per-speaker embedding diagnostics and UMAP/t-SNE visualizations to surface failure modes invisible in aggregate EER.
–Contributed to platform-scale deepfake detection in direct collaboration with Google, analyzing speech of flagged public figures to inform content removal on YouTube.
–All experiments run on UMich Great Lakes HPC via SLURM with multi-GPU PyTorch DDP.

TAI Lab, University of Michigan–Dearborn

Dec 2024 – Aug 2025 · Dearborn, MI

Research Assistant — Computer Vision & ML Systems

Advisor: Prof. Xiao Zhang

–Led a 5-member team that built and published ViKey at IEEE MASS 2025: a $0.20 optical authentication system using polarized birefringence patterns as unclonable physical keys (~80% accuracy at 0.5m), eliminating cloning and replay attack surfaces.
–Designed the computer vision pipeline to distinguish genuine tag optical signatures from environmental reflections using YOLO-based detection and ML classifiers.

Indian Institute of Technology, Indore

Jan 2023 – May 2023 · Indore, India

Research Assistant — Signal Processing + ML

Advisor: Prof. Narendranath Muthumani

–Built a rotation- and scale-invariant star-catalog navigation system using DNNs and normalized cross-correlation, addressing a known fragility of template-matching approaches.
–Reduced RFI contamination in GMRT radio telescope interferometric data using K-Means clustering for unsupervised anomaly flagging.

Publications

IEEE MASS 2025

ViKey: Secure Door Access Control Using Passive Visible Light Tags

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 ~80% auth accuracy at 0.5m while eliminating cloning, replay, and privacy attack surfaces.

Paper → Code →

IEEE MASS 2025

Demo: A Passive Optical Tagging Approach for Secure and Revocable Entry Systems

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.

Paper →

IIETA 2024

A Review of EEG Artifact Removal Methods for Brain-Computer Interface Applications

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.

Paper →

Blog

Mar 2023 · Astronomy

Radio Astronomy

How radio telescopes see what eyes can't, revealing pulsars, black holes, and interstellar gas to map the invisible universe.

Read →

Nov 2023 · History of Computing

Most Important Algorithm That Averted a Nuclear Conflict

How an early-warning algorithm and crucial human judgment prevented a Cold War false alarm from escalating into catastrophe.

Read →