Security and Privacy Challenges in Deep Learning Models Hosted on Cloud Platforms
Jhansi Rani Ganapa, Poonam Joshi, T. Amitha, Sandip Rahane, N. Ravinder, Jignesh Jani, Nilesh Vani, Chandreshkumar Vyas · International Journal of Computational and Experimental Science and Engineering · 2025
Deep learning's fast integration into cloud computing services enables businesses to perform scalable AI model training and real-time analysis in diverse sectors. The combination of deep learning with cloud platforms results in important security vulnerabilities that stem from adversarial threats and data breaches as well as model inversion vulnerabilities and unauthorized system intrusions. Data infringement combined with weakened model reliability and non-compliance requirements require cloud AI systems to adopt more robust security controls. Experts analyze security issues facing deep learning models in the cloud through an assessment of attacks which manipulate model inputs, pollute training data and exploit APIs and create insecurity across multiple cloud user environments. The research compares encryption protocols and federated learning capabilities and access control systems and differential privacy features of AWS, Google Cloud, Microsoft Azure, and IBM Cloud. The assessment evaluates regulatory compliance requirements of GDPR HIPAA and CCPA in order to detect security governance gaps for AI systems. Research outcomes show that Amazon Web Services along with Google Cloud deliver excellent encryption features as well as anomaly detection solutions yet Microsoft Azure stands out through its advanced federated learning functions. The security features aimed at AI operations are insufficient in IBM Cloud which demonstrates divergent approaches to security implementation across platforms. Homomorphic encryption and differential privacy have progressed but practical use remains restricted by high operational costs and regulatory uncertainty as well as attacks by adversaries. The distributed AI training method known as federated learning protects against poisoning attacks but still needs improved protection mechanisms to remain secure. The proposed solution for safe and privacy-compliant AI implementation uses a security system that joins sophisticated cryptographic methods with adversarial attack prevention mechanisms along with methods for protectively training AI. Future research needs to improve encryption speeds as well as strengthen federated learning resistance to attacks and create AI-based compliance systems which will address new cybersecurity threats against cloud-based AI platforms