Aayush Kumar

Aayush Kumar

Auto-generated profile, based on publicly available information

Aayush Kumar

University of Illinois Chicago

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
28 January 2022 in Computational Intelligence and Neuroscience
Verified
Journal Article
Open access
PDF file
AYAnju Yadav
Anju Yadav
RSRahul Saxena
Rahul Saxena
AKAayush Kumar
Aayush Kumar
TWTarandeep Singh Walia
Tarandeep Singh Walia
AZAtef Zaguia
Atef Zaguia
SKS. M. Mostafa Kamal
S. M. Mostafa Kamal
Pulmonary fibrosis is a severe chronic lung disease that causes irreversible scarring in the tissues of the lungs, which results in the loss of lung capacity. The Forced Vital Capacity (FVC) of the patient is an interesting measure to investigate this disease to have the prognosis of the disease. This paper proposes a deep learning-based FVC-Net architecture to predict the progression of the disease from the patients computed tomography (CT) scan and the patients metadata. The input to the model combines the image score generated based on the degree of honeycombing for a patient identified based on segmented lung images and the metadata. This input is then fed to a 3-layer net to obtain the final output. The performance of the proposed FVC-Net model is compared with various contemporary state-of-the-art deep learning-based models, which are available on a cohort from the pulmonary fibrosis progression dataset. The model showcased significant improvement in the performance over other models for modified Laplace Log-Likelihood (6.64). Finally, the paper concludes with some prospects to be explored in the proposed study.

3 December 2025 in IEEE Access
Verified
Journal Article
Open access
AKAayush Kumar
Aayush Kumar
Deepfake media pose a significant threat to digital information integrity, necessitating robust detection strategies. We propose HybridNet, a convolutional network for deepfake detection that integrates residual connections (ResNet), depthwise separable convolutions (EfficientNet principles), and Squeeze–and–Excitation (SE) blocks to balance high accuracy with computational efficiency. HybridNet incorporates uncertainty quantification via softmax confidence thresholds and interpretability through Gradient–weighted Class Activation Mapping (Grad–CAM) visualizations. Enhanced by a Hybrid Frequency–Spatial Attention (HFSA) module, it improves generalization across datasets. We evaluate HybridNet on FaceForensics++ (HQ) (96% accuracy), Celeb-DF v2 (97%), and the DFDC-preview split, where performance improves from 89% (baseline) to 93.2% with data augmentation and PGD-based robustness enhancements. Unlike prior detectors that are either black–box CNNs or heavy transformer architectures, HybridNet offers a practical trade–off between interpretability, efficiency, and accuracy. Additional evaluations against transformer–based models, compression robustness, fairness across demographics, adversarial attacks, and deployment feasibility underscore HybridNet’s versatility and reliability. Ablations confirm the contribution of each architectural component. We report confidence intervals and stress–test under codec compression, noise, and blur to substantiate robustness claims.

1 January 2026 in IEEE Access
Verified
Journal Article
Open access
AKAayush Kumar
Aayush Kumar
Hallucinations hinder reliable question answering, especially in resource-constrained deployments where frontier-scale models or retrieval-heavy pipelines may be impractical. We present EdgeJury, a four-stage ensemble framework that improves truthfulness using only small instruction-tuned language models (3B–8B) suitable for serverless edge inference. EdgeJury combines role-specialized generation, anonymized cross-review, chairman synthesis, and claim-level agreement labeling. On TruthfulQA (MC1), EdgeJury reaches 76.2% accuracy (95% CI: 72.8–79.6%), a + 21.4% relative improvement over a single 8B baseline, and outperforms self-consistency and majority voting under transparent compute accounting. On a 200-question adversarial EdgeCases set, it achieves + 48.2% relative gains, while manual error analysis shows an approximately 55% reduction in factual hallucination errors versus the single-model baseline. Beyond end-to-end gains, we analyze why the framework works and when it is worth its cost. Stage 1 agents exhibit non-trivial complementarity, blind synthesis preserves most of the full system’s accuracy while reducing chairman self-preference, structured critique outperforms open-ended critique in orchestration reliability, and scaling/pruning experiments show that most of the gain is obtained by four agents with substantial latency recovery available through early exit. Deployed on Cloudflare Workers AI, EdgeJury achieves 8.4 s median end-to-end latency, showing that coordinated small-model ensembles can improve truthfulness on misconception-heavy QA benchmarks without external retrieval or proprietary large-model APIs.