Broader Research Interests

Beyond our current projects, we are interested in expanding into these adjacent research areas. We welcome collaborations and project proposals from researchers working in these domains.

Future Research Directions

Adversarial Robustness Metrics for Synthetic Media Detection

Benchmarking and evaluation frameworks for deepfake detection systems

Development of standardized evaluation protocols and robustness metrics for detecting AI-generated media. This includes adversarial perturbation analysis, cross-domain generalization benchmarks, and detection-evasion trade-off quantification under realistic deployment conditions.

Deepfake detection Adversarial robustness Benchmark design Forensic analysis

Physics-Informed Generative Models for CAD and AEC Applications

Constraint-aware image synthesis for architecture, engineering, and construction

Integration of physical constraints (structural loads, material properties, manufacturing tolerances) into generative architectures for producing design candidates that satisfy real-world engineering requirements. Applications span parametric facade generation, structural topology optimization, and fabrication-aware geometry synthesis.

Physics-informed ML CAD generation Structural constraints Parametric design

Semantic Grounding and Prompt-Output Alignment in Diffusion Models

Measuring and enforcing text-to-image fidelity under compositional prompts

Research into quantifiable alignment metrics between natural language prompts and generated visual outputs, with focus on compositional semantics, attribute binding accuracy, spatial relationship preservation, and negation handling. Includes development of automated evaluation pipelines and alignment-aware fine-tuning strategies.

Prompt alignment Compositional generation Semantic grounding T2I evaluation

Distributed Inference and Privacy-Preserving Vision Systems

Federated learning and edge deployment for resource-constrained computer vision

Exploration of federated learning protocols for vision models that preserve data privacy while enabling collaborative model improvement. Research includes communication-efficient gradient compression, differential privacy guarantees, heterogeneous device optimization, and latency-bounded inference for edge deployment scenarios.

Federated learning Edge deployment Privacy-preserving ML Model compression

Open Call for Collaborations

Have a project idea that aligns with our research interests? We are open to collaborating with researchers, students, and practitioners who want to pitch projects in these areas—and beyond, as long as they fall within the broader scope of constrained image synthesis and safety-aligned generative models.

To propose a collaboration: Send a brief project description (1-2 paragraphs) and your background to ahujaangadsingh@gmail.com