Ratchet-TBC: Self-healing Tweakable Block Ciphers for Post-compromise
Secure Messaging
Jaiswal, A. K. & Krishna, V.
Journal of Cryptology
2025
This paper introduces Ratchet-TBC, a novel cryptographic construction that
integrates self-healing mechanisms into tweakable block ciphers for secure
messaging protocols. Our approach enables automatic key recovery after
compromise, ensuring both forward and backward secrecy in end-to-end encrypted
communications. We provide formal security proofs and demonstrate practical
efficiency through implementation benchmarks.
Keywords: Tweakable Block Ciphers, Post-Compromise Security,
Ratcheting, Secure Messaging, Forward Secrecy
Streaming Stackelberg Security for Cloud-Native Microservices via Graph
Neural Bilevel Optimization
Jaiswal, A. K. & Kumar, S.
Journal of Network and Computer Applications
2025
We present a game-theoretic framework for securing cloud-native microservices
architectures using Stackelberg equilibrium concepts. By modeling the
attacker-defender interaction as a bilevel optimization problem and leveraging
graph neural networks to capture service dependencies, our approach enables
real-time security policy adaptation in dynamic cloud environments. Experimental
results demonstrate significant improvements in threat mitigation and resource
allocation efficiency.
Keywords: Microservices Security, Stackelberg Games, Graph Neural
Networks, Cloud Computing, Bilevel Optimization
Integrating Human Feedback in Caller ID Spoofing Defenses: A Hybrid Approach
for Real-time VoIP Security and Adversarial Robustness
Krishna, V. & Jaiswal, A. K.
Computer Networks
2025
This research addresses the growing threat of caller ID spoofing in VoIP systems
by proposing a hybrid defense mechanism that combines machine learning-based
anomaly detection with human-in-the-loop feedback. Our system achieves real-time
threat identification while maintaining adversarial robustness through continuous
learning from user reports. We validate the approach using real-world VoIP traffic
datasets and demonstrate superior performance compared to purely automated solutions.
Keywords: VoIP Security, Caller ID Spoofing, Human-in-the-Loop,
Adversarial Robustness, Machine Learning
Beyond Accuracy: A Mechanistic and Ecological Failure Analysis of Zero-Shot
Large Multimodal Models for Fine-Grained Crop Disease Diagnosis
Krishna, V., Jaiswal, A. K., & Mahajan, S.
Computers and Electronics in Agriculture
2025
We conduct a comprehensive analysis of large multimodal models (LMMs) for
zero-shot crop disease diagnosis, revealing critical failures in fine-grained
classification tasks. Through mechanistic interpretability techniques and
ecological validation, we identify specific failure modes related to visual
attention, reasoning biases, and domain shift. Our findings provide actionable
insights for improving AI-based agricultural diagnostic systems.
Keywords: Large Multimodal Models, Zero-Shot Learning, Crop Disease
Diagnosis, Failure Analysis, Agricultural AI
Privacy-preserving Hybrid Multi-domain Image Watermarking with Fully Homomorphic
Encryption and Content-adaptive Embedding under Realistic Attack Scenarios
Krishna, V., Jaiswal, A. K., & Singh, N. P.
Manuscript in Preparation
2025
We propose a novel privacy-preserving watermarking scheme that operates in
encrypted domain using fully homomorphic encryption (FHE). By combining spatial
and frequency domain embedding with content-adaptive selection, our method
achieves robust watermark detection under realistic attack scenarios including
JPEG compression, geometric transformations, and noise addition, all while
maintaining complete privacy of the host image.
Keywords: Digital Watermarking, Homomorphic Encryption, Privacy
Preservation, Multi-domain Embedding, Adversarial Attacks
Emotion-aligned Cross-modal Fusion for Speech Emotion Recognition: Contrastive
Learning, Low-rank Attention, Gender-invariant Modelling, and Zero-shot Generalization
Krishna, V., Kumar, A., & Jaiswal, A. K.
Computer Speech & Language
2025
This paper introduces an emotion-aligned framework for speech emotion recognition
that integrates contrastive learning, low-rank attention mechanisms, and
gender-invariant modeling. Our approach enables effective cross-modal fusion of
acoustic and linguistic features while achieving zero-shot generalization to
unseen emotional categories. Extensive experiments on benchmark datasets demonstrate
state-of-the-art performance and robustness across diverse speakers and recording
conditions.
Keywords: Speech Emotion Recognition, Cross-modal Fusion, Contrastive
Learning, Zero-shot Learning, Gender Invariance