DCASS

Project highlights

• Zero-modification carrier strategy designed to resist classical content-level steganalysis.
• Cross-modal retrieval engine supporting text, image, and audio carriers from a shared semantic search interface.
• GAN + RL stealth stack that learns human-like dispatch timing under throughput and suspicion constraints.
• Dockerized sender-receiver simulation for controlled end-to-end experiments and reproducible evaluation.

What it is

DCASS (Dynamic Context-Aware Semantic Steganography) is a research-grade covert communication system that encodes intent through semantically matched media retrieval instead of editing carrier files. It combines multi-modal embeddings, vector search, behavior-aware scheduling, and adversarial evaluation to test whether hidden communication can remain statistically indistinguishable from normal content sharing.

Problem it solves

Traditional steganography alters pixels, bytes, or waveforms, which can leave detectable artifacts under modern steganalysis. DCASS addresses that by keeping carriers untouched and encoding via semantic curation, then reducing traffic-level detectability with timing and channel-behavior controls.

How it works

• Ingest and index text, image, and audio corpora into a unified semantic retrieval layer using CLIP, CLAP, and sentence-level embeddings with FAISS-backed nearest-neighbor search.
• Chunk each message into semantic units, generate query vectors, and map each unit to naturally occurring carriers instead of embedding bits into files.
• Apply dynamic context keying so carrier mappings shift with time and context inputs, reducing static pattern reuse across sessions.
• Route selected carrier sequences through the stealth layer, where GAN-based temporal generation and PPO-style policy optimization shape transmission behavior.
• Evaluate output through an adversarial warden pipeline and benchmark scripts that score detectability, throughput, and reconstruction quality.

Key capabilities

• Zero-modification carrier strategy designed to resist classical content-level steganalysis.
• Cross-modal retrieval engine supporting text, image, and audio carriers from a shared semantic search interface.
• GAN + RL stealth stack that learns human-like dispatch timing under throughput and suspicion constraints.
• Dockerized sender-receiver simulation for controlled end-to-end experiments and reproducible evaluation.
• Extensible docs and benchmark flow for architecture, sequence behavior, and corpus expansion experiments.

Impact and outcomes

• Demonstrates a practical transition from bit-level hiding to meaning-level covert encoding.
• Provides a reproducible research baseline for adversarial traffic-aware steganography experiments.
• Establishes a modular foundation for future real-channel deployment and comparative steganalysis studies.