publications

FreqDINO: Multi-Modal Deepfake Detection with Band-Conditioned Cross-Attention

Arush Gumber, et al.

CVPR 2026: Manuscript under review (First Author)

Proposed FreqDINO, a multi-domain deepfake detector fusing semantic, frequency, and noise cues via a novel Band-Conditioned Phased Cross-Attention (BC-PCA). Achieved 98% F1 vs. 93.6% baseline with interpretable CLIP/DINO-based U-Net decoder.

TOAM-YOLO: A Lightweight Tiny Object-Aware Multi-Expert Framework

Arush Gumber, et al.

AAAI 2026: Manuscript under review (First Author)

Tiny object detection with 4M parameters achieving SOTA on 5 benchmarks. Novel TOA-MoE module with BiFPN fusion and CARAFE upsampling for efficient multi-scale feature processing.

DentalNet: Geometric Aware Multi-View Transformer for Dental 3D Scan Analysis

Arush Gumber, et al.

NeurIPS 2025 Imageomics Workshop: Accepted (First Author)
AAAI 2026: Under review

Multi-modal 2D/3D fusion achieving 67% F1-score. Cross-attention mechanism integrates 2D intraoral views with 3D point cloud representations for IOTN-DHC classification.

Unified Framework for Underwater Acoustic Event Detection

Arush Gumber, et al.

TMLR: In preparation

Harmonized datasets for marine soundscape monitoring. Standardized preprocessing pipeline with deep learning architecture for biodiversity assessment and vessel tracking.

MSCAN: Multistage Spinal Canal Stenosis Grading Using Cross-Attention

Arush Gumber, et al.

IEEE/CVF CVPR 2025 Demo Track: Accepted

Deep learning framework for automated MRI-based spinal stenosis grading achieving 0.971 AUROC with YOLO detection and multi-view cross-attention.

SocialDF: Benchmark Dataset and Detection Model for Deepfake Content

Arush Gumber, et al.

ACM ICMR 2025: MAD Workshop: Accepted

Real-world benchmark with multi-agent LLM framework outperforming SOTA lip-sync methods through fact-checking integration.