Projects

Technical Implementation

• Dynamic Analysis: AST instrumentation and object serialization (Pickle) to trace/compare variable states across runs using VarDeterminismChecker.
• Static Taint Analysis: Integrated with Meta's Pysa to trace nondeterminism propagation from sources to outputs.
• SMT Encoding: Encodes algorithm logic into Z3 SMT format to check for satisfiability.

Key Features

• ND Kernel Extraction: Isolates the minimal code subset responsible for propagating nondeterminism.
• Taint Dependency Mapping: Generates call and dependency graphs to visualize data flow.
• Over-approximation Filtering: Prunes variables (e.g., clustering labels) that don't affect final semantics.

Core Capabilities

• Developed CMC and TAGD: Engineered dual algorithmic frameworks to actively steer model generalization and obfuscate capabilities post-convergence.
• Controlled Misclassification (CMC): Nudges converged models out of overfitted, sharp loss regions by introducing minimal final-layer weight perturbations to flip targeted labels, boosting test accuracy by up to 2.8% after resuming training.
• Model Obfuscation (TAGD): Implements Training Accuracy-preserving Generalization Degradation to deliberately lower prediction confidence and test performance on unseen data while keeping training labels and accuracy completely intact.

Technical Architecture

• Formulated and solved exact optimization constraints using the Gurobi Optimizer to compute precise weight and bias updates without initial retraining.
• Surgically modifies only the final layer of the network, ensuring scalable optimization runtime independent of the overall network depth.
• Evaluated and validated across diverse network architectures including ResNet18, ResNet50, WideResNet, and Network-in-Network (NIN).
• Tested performance across 10 multiclass vision benchmarks (such as CIFAR10, SVHN, MNIST, and Caltech101) and 5 binary tabular datasets.

Core Approach

Feedback-Directed Search: Uses a gradient descent-like approach (univariate and bivariate search) to explore a toolkit's hyperparameter space.

Zero Code Modifications: Operates entirely in a black-box manner, reducing nondeterminism and inconsistency without requiring access to source code.

Core Capabilities

• Supported Toolkits: Optimizes popular algorithms across Scikit-learn, MATLAB, and R implementations.
• Targeted Algorithms: Handles anomaly detection algorithms, including Isolation Forest (IF), Robust Covariance (RobCov), Local Outlier Factor (LOF), and One-Class SVM (OCSVM).

Core Goal

Algorithm-independent acceleration through parallelization and ensemble evaluation methods.

Architecture

• Optimized for Scikit-Learn integration.
• Focuses on reducing execution time without compromising clustering quality.