Kasra Ahmadi

Kasra Ahmadi

Ph.D. Candidate, Computer Science & Engineering — University of South Florida

ENB 323, 3820 USF Alumni Drive, Tampa, FL 33620

Email Scholar GitHub LinkedIn YouTube Resume Full CV

About

I am a Ph.D. candidate in the Department of Computer Science and Engineering at the University of South Florida, advised by Dr. Mehran Mozaffari Kermani and co-advised by Dr. Rouzbeh Behnia.

My research is in machine learning security and post-quantum cryptography. I design algorithm-level error-detection and fault-attack countermeasures for PQC schemes (Kyber, Dilithium) and classical ECC/elliptic-curve hardware, and study side-channel and fault-injection resilience for cryptographic and DNN inference pipelines. On the ML side, I work on privacy-preserving training (differentially private federated learning) and adversarial robustness (evasion attacks against RAG retrieval). This work won a Best Paper Award at IEEE S&P 2025 Workshops.

I'm currently an ML Research Engineer Intern at RealmLabs, building production PII-detection and model-routing systems for Claude Code — bridging side-channel and adversarial-ML research with real-world ML security engineering.

Experience

ML Research Engineer Intern · RealmLabs, Sunnyvale, US
May 2026 – Present
  • Developed PII privacy and security workflows using frozen Qwen 3.5 and Llama 3.1 models, achieving a 3% F1 score improvement over ModernBERT by disabling causal masking to capture full bidirectional context.
  • Developing probe models on top of decoder hidden states to detect PII at the token level.
  • Working on cost optimization and model routing for Claude Code by classifying session intent and routing sessions to the appropriate Claude model.
  • Integrating vLLM with sparse autoencoders (SAEs) to capture token-level signals during generation in a single pass.
Research Assistant · University of South Florida, Tampa, US
Jan 2022 – Present
  • Designed a human-in-the-loop, differentially private federated learning framework for fine-tuning LLMs (BERT) on memory-constrained devices, solving the fixed-memory limitation of existing DP mechanisms; won Best Paper Award at IEEE S&P 2025 Workshops (HMI-SA).
  • Built MAED, an algorithm-level runtime error-detection framework using mathematical identities to catch fault-injection attacks and hardware faults on DNN activation functions (ReLU, sigmoid, tanh), hardening embedded ML inference pipelines against adversarial and natural faults.
  • Designed algorithm-level error detection and side-channel/fault-attack countermeasures for NIST-standardized post-quantum cryptosystems (CRYSTALS-Kyber, Dilithium) and classical public-key schemes (ECC/ECSM, Montgomery Ladder), benchmarked on FPGAs, ARM, and embedded Linux; published in IEEE TVLSI, IEEE TCAD, and ACM TECS.
  • Authored 10+ peer-reviewed papers spanning ML security, privacy-preserving AI, and hardware security under an NSF-funded research program (Award #1801488); serve as a peer reviewer for IEEE TVLSI, TCAS-I, and ACM TECS.
  • Teaching assistant for Deep Learning in Computer Vision, Cryptography, Operating Systems, Computer Architecture, Network Lab, and System Design.
AI Engineer Intern · TD Synnex, Clearwater, US
May 2025 – Aug 2025
  • Designed and implemented cloud-native AI workflows combining LLMs, RAG, and multi-agent orchestration.
Software Engineer Intern · AgWise (TransparencyWise), St. Petersburg, US
May 2024 – Aug 2024
  • Built an AWS-based ETL pipeline using Lambda, Glue, and S3 to process laboratory nutrient data for crop recommendation models, supporting nutrient recommendations tailored to various growth stages of corn and soybeans.
  • Implemented an event-driven architecture utilizing Lambda functions, Step Functions, Event Bridge, SES, and API Gateway to promote loose coupling and scalability.
ML Research Engineer · Paar Lift, Tehran, IR
Jan 2018 – Apr 2020
  • Analyzed optimal floor levels for elevators at specific times to reduce passenger wait times using machine learning, such as logistic regression and KNN.
  • Established a connection between Raspberry Pi embedded boards and elevators through the CAN bus protocol for real-time data transfer to a Linux-powered IoT system.
  • Built ETL pipelines using Apache Airflow to extract, ingest, and load elevator traffic data to an OLAP store.
  • Reduced passenger wait times by 27% across deployments in 100+ commercial buildings.

Education

Ph.D. in Computer Science · University of South Florida, US
Jan 2022 – Present
CGPA: 3.93/4
MSc in Information Technology Engineering · Amirkabir University of Technology, Iran
Sep 2018 – Aug 2020
BSc in Computer Science · Isfahan University of Technology, Iran
Sep 2012 – Jul 2017

Publications

  1. An Interactive Framework for Implementing Privacy-Preserving Federated Learning: Experiments on Large Language Models Kasra Ahmadi, Rouzbeh Behnia, Reza Ebrahimi, Mehran Mozaffari Kermani, Jeremiah Birrell, Jason Pacheco, Attila A. Yavuz In 2025 IEEE Security and Privacy Workshops (SPW), San Francisco, CA, USA. Best Paper Award.
  2. Efficient Algorithm Level Error Detection for Number-Theoretic Transform Assessed on FPGAs Kasra Ahmadi, Saeed Aghapour, Mehran Mozaffari Kermani, Reza Azarderakhsh ACM Transactions on Embedded Computing Systems, 2025.
  3. MAED: Mathematical Activation Error Detection for Mitigating Physical Fault Attacks in DNN Inference Kasra Ahmadi In submission.
  4. PUF-Kyber: Design of a PUF-Based Kyber Architecture Benchmarked on Diverse ARM Processors Saeed Aghapour, Kasra Ahmadi, Mila Anastasova, Mehran Mozaffari Kermani, Reza Azarderakhsh IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2024.
  5. Efficient Error Detection Schemes for ECSM Window Method Benchmarked on FPGAs Kasra Ahmadi, Saeed Aghapour, Mehran Mozaffari Kermani, Reza Azarderakhsh IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 32, no. 3, pp. 592-596, March 2024.
  6. PUF-Dilithium: Design of a PUF-Based Dilithium Architecture Benchmarked on ARM Processors Saeed Aghapour, Kasra Ahmadi, Mehran Mozaffari Kermani, Mila Anastasova, Reza Azarderakhsh ACM Transactions on Embedded Computing Systems, vol. 24, no. 2.
  7. Efficient Error Detection Cryptographic Architectures Benchmarked on FPGAs for Montgomery Ladder Kasra Ahmadi, Saeed Aghapour, Mehran Mozaffari Kermani, Reza Azarderakhsh IEEE Transactions on Very Large Scale Integration (VLSI) Systems, early access, 2024.
  8. A P2P File Sharing Market Based on Blockchain and IPFS with Dispute Resolution Mechanism Kasra Ahmadi, Molud Esmaili, Siavash Khorsandi IEEE Int'l Conf. on Artificial Intelligence, Blockchain, and Internet of Things (AIBThings), Mount Pleasant, MI, pp. 1-5, 2023.
  9. [Preprint] Error Detection Schemes for τ NAF Conversion within Koblitz Curves Benchmarked on Various ARM Processors Kasra Ahmadi, Saeed Aghapour, Mehran Mozaffari Kermani, Reza Azarderakhsh TechRxiv.
  10. [Preprint] Efficient Fault Detection Architectures for Modular Exponentiation Targeting Cryptographic Applications Benchmarked on FPGAs Saeed Aghapour, Kasra Ahmadi, Mehran Mozaffari Kermani, Reza Azarderakhsh arXiv preprint arXiv:2402.18033
  11. [Preprint] Envisioning the Future of Cyber Security in Post-Quantum Era: A Survey on PQ Standardization, Applications, Challenges and Opportunities Saleh Darzi, Kasra Ahmadi, Saeed Aghapour, Attila Altay Yavuz, Mehran Mozaffari Kermani arXiv:2310.12037, 2023.

Projects

Federated LLM Fine-Tuning with Differential Privacy

Human-in-the-loop framework for privacy-utility trade-off control in federated fine-tuning of BERT-based models using Flower, DP-SGD, and LoRA. Best Paper Award, IEEE S&P 2025 Workshops.

MAED: Error Detection for Fault Attacks on DNN Inference

Algorithm-level error-detection methods using mathematical identities to catch fault-injection attacks and hardware faults on DNN activation functions (ReLU, sigmoid, tanh) for safety- and security-critical inference deployments.

Fault-Attack Simulation for Classical and Post-Quantum Cryptography

Simulated fault-injection attacks on Kyber, Dilithium, and ECC, then proposed lightweight algorithm-level countermeasures with near-100% error detection, benchmarked on FPGAs and ARM.

Gradient-Based Adversarial Attacks Against RAG Systems

Evasion attacks that perturb vector embeddings to degrade RAG retrieval quality, answer reliability, and downstream system accuracy.

PII-360

Open-source Chrome extension that detects personally identifiable information in images and PDFs using ONNX models running locally on-device.

Certifications

Service