Automated Classification of E-Commerce Packaging for Open Box Delivery

Shiyamaladevi R S, Jeetashree Aparajeeta. Automated Classification of E-Commerce Packaging for Open Box Delivery: A Transfer Learning Solution. TechRxiv. November 10, 2025. DOI: 10.36227/techrxiv.176281084.44193627/v1

• Developed a transfer learning–based image classification pipeline to detect poor-quality packaging images (Blur/Dark/Random, Primary/Secondary Package).
• Achieved 89% overall accuracy and 96% recall for low-quality images using fine-tuned CNN architectures (ResNet50 + augmentation).
• Deployed an end-to-end automated workflow integrated with Google Sheets—allowing non-technical operators to run real-time OBD quality checks without coding.
• Reduced manual inspection overhead by 60%, improving packaging integrity assurance across last-mile delivery systems.

Swara-Net: A Deep Learning Baseline for Under-Represented Carnatic Ragas

Selected for Dec IEEE INDICON’2025

Shiyamaladevi R S. Swara-Net: A Deep Learning Baseline and Curated Dataset for Five Under-Represented Carnatic Ragas. TechRxiv. October 31, 2025. (Selected for Dec IEEEINDICON’2025) DOI: 10.36227/techrxiv.176281084.44193627/v1

• Curated a diverse dataset of 638 audio snippets across five ragas (Shankarabharanam, Revathi, Abheri, Brindavanasaranga, Mayamalavagowla).
• Designed Swara-Net, a 3-layer Deep Neural Network achieving 99.0% classification accuracy, outperforming SVM by 42%.
• Conducted detailed musicological error analysis of the model’s misclassification patterns and published the open-source dataset for reproducibility.
• Pioneered the first computational baseline for these five ragas, contributing to computational ethnomusicology research in Indian classical music.

Security Auditing & Malicious Client Blocking in Federated EEG Seizure Detection

Under Review – Frontiers in Neuroinformatics

Shiyamaladevi R S, Swaroop S Kaimal, Jeetashree Aparajeeta. Security Auditing and Ultra-Strict Malicious Client Blocking in Federated EEG Seizure Detection. TechRxiv. (Under review at Frontiers)

• Engineered a three-tier attestation-based federated learning framework integrating HMAC-SHA256 for secure client verification.
• Achieved 97.5% accuracy in no-attack and 95.2% accuracy under 30% malicious clients, with 98.7% attack detection rate.
• Optimized lightweight CNN models for Raspberry Pi deployment, achieving 18 ms inference latency with 99% validation accuracy.
• Established a HIPAA/GDPR-compliant architecture enabling secure, privacy-preserving clinical AI for distributed seizure detection.

Alleviating Muscle Fatigue in Robotic Prosthetic Arms via Multi-Sensor Fusion

Completed - Published in IEEE

Manimaran, P., Ashwin, J. B., Swaroop, S. K., & Shiyamaladevi, R. S. (2024, December). Alleviating Muscle Fatigue in Robotic Prosthetic Arms Through Multi-Sensor Fusion and Adaptive Real-Time Grip Assistance. In 2024 International Conference on Communication, Computing, Smart Materials and Devices (ICCCSMD) (pp. 1-12). IEEE. DOI: 10.1109/ICCCSMD63546.2024.11015135

• Designed a multi-sensor adaptive control system integrating EMG signals, computer vision, and pressure feedback to optimize grip force.
• Proposed a real-time fatigue detection algorithm reducing muscle strain and improving prosthetic control comfort.
• Validated feasibility for hardware-level integration with existing prosthetic arms, enhancing long-term usability and efficiency.