About This Research Area

Efficient AI is not enough if it cannot be trusted. We build hardware for privacy-preserving computation, lightweight intrusion detection, and cryptography on resource-constrained devices.

Privacy-preserving AI hardware

About the Project

Table comparing logic cells, accuracy and training throughput for differential privacy accelerator designs.
Resource and throughput results on ZCU102.

Differential privacy protects training data by adding calibrated noise — but generating that noise in software is slow, and moving it around costs memory bandwidth.

  • Gaussian noise generated directly in hardware by exploiting circuit metastability
  • Paired with precision-aware multipliers
  • Privacy comes at almost no throughput cost

Haider, Kim, Zhang, Arias-Garcia, Lee and Ko, IEEE APCCAS, 2025.

LightIDS: lightweight intrusion detection

About the Project

Grouped bar chart comparing precision, recall and F1-score of ten intrusion detection methods.
Precision, recall and F1-score against nine published systems.

Network intrusion detection usually means large models running on servers. LightIDS is an abstracted, high-precision deep neural network small enough to run in constrained settings.

  • Small enough for edge and embedded deployment
  • Competitive with much heavier published systems
  • Evaluated on the CIC-IDS2017 benchmark

Fard, Soltani, Jahangir and Ko, The Journal of Supercomputing, 2025.

Cryptographic hardware

About the Project

The group has a long line of work on cryptographic circuits, aimed at devices where area, power and throughput are all constrained at once.

  • High-throughput and area-efficient AES implementations
  • Nano-AES for IoT devices where area and power are scarce
  • Post-quantum cryptosystems for resource-constrained hardware