AI-Guided Adaptive Compression for Secure and Efficient Web Resource Delivery

Modern web applications increasingly rely on large volumes of client-side static resources, making delivery-layer efficiency a critical concern for both performance and cybersecurity resilience. Although transport-level compression techniques such as Gzip and Brotli are widely supported, they are typically applied statically without considering runtime context such as asset characteristics, network conditions, or device constraints. This static approach limits performance gains and increases exposure to network-based threats.This paper presents an AI-guided adaptive compression framework that dynamically selects transport-level compression strategies for web resource delivery. We instantiate the decision layer with a lightweight tree-based classifier (Random Forest) that predicts the compression choice per resource from runtime features such as Content-Type, payload size bucket, observed RTT, and concurrency context. The policy minimizes a cost function that balances expected transfer-time reduction against CPU overhead. A controlled experimental setup models representative web workloads composed of heterogeneous static assets delivered under parallel and sequential loading strategies.Results show that adaptive compression outperforms static baselines across the evaluated workload and network profiles. While individual asset compression yields moderate gains, cumulative effects across concurrent loading achieve aggregate transfer reductions exceeding ten percent, leading to measurable improvements in First Contentful Paint and Largest Contentful Paint. Beyond performance, adaptive compression reduces network exposure duration and can lower the effectiveness of timing-based traffic analysis by shortening and smoothing observable transfer windows. These findings position adaptive compression as a practical, under-explored control surface for AI-enabled, security-aware web infrastructure.