Aniket Roy

Wavelet-domain supervision for 3D Gaussian Splatting, combining multi-scale coefficient alignment, kurtosis concentration, and cross-band covariance penalties to improve rendering quality, especially from sparse views.

An overview of my doctoral research on resource-efficient generative AI: uncertainty-guided mixup for few-shot learning, diffusion models grounded in natural image statistics, and parameter-efficient low-rank adapters for personalized synthesis.

Content-style personalization of diffusion models via cycle-consistent training and layer-wise diffusion priors.

A training-free method for composing multiple LoRA adapters using frequency-domain guidance during diffusion sampling.

A kurtosis-based loss grounded in natural image statistics that improves the perceptual quality of diffusion-generated images.

Controlled diffusion conditioned on target-domain context and style turns synthetic renders into realistic training images, raising VisDA target accuracy from 90.3% to 91.8%.

Uses diffusion-generated data to reduce demographic bias in deep classifiers.

Generalizing action recognition from ground-level video to aerial viewpoints.

Semantic data augmentation for low-data regimes using pretrained captioning and text-to-image diffusion models.

Generates hard latent positives to improve contrastive self-supervised learning of skeleton-based action encoders.

A differentiable regularizer that lower-bounds the distance of data points to the decision boundary, improving certified robustness.

Improves few-shot classification by generating hard mixup samples that sharpen decision boundaries.

Leverages text-to-image diffusion models to synthesize and align training data for few-shot recognition.

Optimal-transport-based fusion of video, audio, and text for detecting sarcasm and humor.

A benchmark dataset for brightness illusions, with an analysis of how diffusion models perceive illusory regions.

An adversarial framework that suppresses protected attributes in face descriptors to reduce demographic bias.

A knowledge-distillation framework for reducing demographic bias in face recognition.

A monograph covering the theory and practice of digital image forensics.

A theoretical analysis of prediction-error-expansion reversible watermarking toward optimal embedding.

A recurrent attention model for distinguishing computer-generated from natural images.

DCT residual features for detecting filtering forgery and splicing in JPEG images.

Camera source identification using DCT residue features with an ensemble classifier.

Rotated local binary pattern features to separate copy-move forgeries from visually similar genuine objects.

Detection and localization of JPEG forgeries based on noise correlation.

Theoretical analysis of embedding distortion in prediction-error-expansion reversible watermarking.

A human-visual-system-inspired watermarking scheme with a robustness analysis.

A reversible color image watermarking scheme with an information-theoretic analysis.