Research and Selected Projects

My work connects trustworthy machine learning, optimization, and statistical inference. These projects summarize methods and documented results; Experience describes my responsibilities and appointments.

Scalable Neural-Network Robustness Verification

• Developed sparse second-order cone programming (SOCP) relaxations to strengthen certified robustness guarantees for ReLU networks, using selective McCormick lifting and local 2 × 2 positive-semidefinite constraints.
• Designed CORE, a class-wise verification cascade that matched SDP-certified robust accuracy while reducing average verification time by up to 86.30% on a fully connected MNIST benchmark.
• Evaluated fully connected and convolutional networks on MNIST and CIFAR-10 against LP, SDP, and α,β-CROWN baselines using certified robust accuracy and runtime.

CoRe-CROWN: Conic-Regularized Certified Robust Training

• Developed certified robust training that combines a Dual-CROWN objective with sparse SOCP-derived regularization applied during training.
• Designed influence-based selection of unstable ReLU units and local lifted constraints using McCormick envelopes and 2 × 2 positive-semidefinite constraints.
• Evaluated clean accuracy, adversarial robustness, and complete certified accuracy to examine the effect of the training regularizer.

Sequential Distribution-Shift Detection — DI-SCUSUM

• Co-developed a training-free method for streaming change detection using closed-form scores across Gaussian smoothing scales.
• Established detection-delay and false-alarm guarantees linked to Kullback–Leibler divergence under empirical models.
• Evaluated Gaussian simulations, MNIST, and Oxford-IIIT Pet at matched false-alarm levels. In a calibrated anisotropic Gaussian experiment, achieved approximately 91% lower detection delay than score-based CUSUM, nearly matching likelihood-ratio CUSUM.

Equal contribution: Arman Adibi and Mohammadreza Maleki. Collaboration with Sanjeev Kulkarni and H. Vincent Poor at Princeton University.

Secure Machine Learning — UN PET Lab Hackathon

Secure Machine Learning course project, December 2022.

• Built a prediction workflow for hidden training labels using privacy-budgeted queries and SmartNoise.
• Applied missingness screening, histogram analysis, and correlation analysis; compared linear regression, random forest classification, and logistic regression.
• Obtained a 0.6934 competition score with logistic regression at total epsilon expenditure of 1.85. This is the reported competition score, not a classification-accuracy percentage.
• Extended the study to heart-disease prediction: reported decision-tree accuracy was 57% without differential privacy and 52% with epsilon = 0.1 in the documented experiment.

Code, methods, reported results, and limitations

Cascading and Ensemble Robustness

• Developed staged verification that applies stronger methods only to unresolved cases and evaluates certificate-quality/runtime trade-offs.
• Contributed to CEAR, an ensemble robustness framework evaluated on MNIST, CIFAR-10, and Tiny ImageNet.

See the SaTML and Canadian AI publications for paper summaries and downloads.

Earlier Engineering Projects

• Master’s thesis — Phase Array Inductive Wireless Power Transfer for Moving Load: University of Tehran; supervised by Prof. Shahin Jafarabadi Ashtiani. Research involved mathematical modeling, optimization, and simulation of inductive wireless power-transfer systems.
• Bachelor’s capstone — Basics and Applications of Digital Controlled Oscillators: Isfahan University of Technology; supervised by Prof. Masood Omoomi.
• X-band microwave amplifier: Bachelor’s elective project on amplifier design and simulation; supervised by Prof. Abolghasem Zeidaabadi Nezhad.
• Deep Brain Stimulator national project: Contributed to the design, optimization, and evaluation of Class-E/D amplifiers, power and data coils, and rectifiers.

For a consolidated summary, see my CV or download the two-page PDF.