Skip to content
Karamveer Singh Sidhu

Research

Research and publications

Alongside my payments work, I research ML security: adversarial robustness in federated learning and privacy for LLM systems. Here are my MSc thesis, research code, and publications.

01MSc thesis

Causal Structure Learning for Adversarial Robustness in Federated Learning

University of Northern British Columbia · MSc Computer Science · 2024 to 2026

CNN-based intrusion detection systems are highly vulnerable to adversarial evasion attacks. I tested causal structure learning on a federated, CNN-based intrusion detection system. It raised the model's robustness under PGD from 41% to 89.85% at epsilon 0.20, with under 1% clean-accuracy loss and no adversarial training.

The model used a learnable sparse 64x64 causal matrix, aggregated across clients with FedAvg. I evaluated it on a 4-client non-IID federation built with PyTorch and Flower, under FGSM and PGD attacks. The training data came from my pipeline, which merges four network intrusion detection datasets into a 109M-row unified corpus.

Read the thesis·causal-fl-ids-defense·um-nids-merge-pipeline

Causal structure learning for adversarial robustness in federated learning

Karamveer Singh Sidhu

MSc · Computer Science
University of Northern
British Columbia

2026

02Award

Mitacs Globalink Research Award

City, University of London · Jan to Mar 2026

During this 12-week research internship at City, University of London, I researched and implemented the core causal structure learning algorithm behind my thesis result.

03Publications

ITU · VOL. 7, ISSUE 22026

FLAccShield: Federated learning with accuracy-weighted aggregation for DDoS detection

Md Mahibul Hasan, Aymen Ben Said, Karamveer Singh Sidhu, Nashid Shahriar, Sajal Saha

ITU Journal on Future and Evolving Technologies · Volume 7, Issue 2

We proposed FLAccShield, a privacy-preserving federated learning framework for DDoS detection across 5G network slices. In supervised non-IID tests, it outperformed FedAvg and FedProx by approximately 3% in F1 score.

arXiv · 2601.100042026

SoK: Privacy-aware LLM in Healthcare: Threat Model, Privacy Techniques, Challenges and Recommendations

Mohoshin Ara Tahera, Karamveer Singh Sidhu, Shuvalaxmi Dass, Sajal Saha

arXiv preprint

Large Language Models (LLMs) are increasingly adopted in healthcare to support clinical decision-making, summarize electronic health records (EHRs), and enhance patient care. However, this integration introduces significant privacy and security challenges, driven by the sensitivity of clinical data and the high-stakes nature of medical workflows. This Systematization of Knowledge (SoK) examines the evolving threat landscape across the three core LLM phases: data preprocessing, fine-tuning, and inference within realistic healthcare settings. We present a detailed threat model that characterizes adversaries, capabilities, and attack surfaces at each phase, and we systematize how existing privacy-preserving techniques attempt to mitigate these vulnerabilities. We conclude with phase-aware recommendations and future research directions aimed at strengthening privacy guarantees for LLMs in regulated environments.