Home / Projects / Deceptive-Net
Case Study · Academic Capstone · Jan – Jun 2026
Deceptive-Net — Honey-Token Generative Framework
Attackers reconnoitre networks and hunt for credentials long before defenders notice. Deceptive-Net turns that reconnaissance against them: an AI-driven defensive system that uses a Wasserstein GAN (WGAN-GP) to generate hyper-realistic honey-tokens and deceptive network artefacts, so the first thing an attacker touches is a trap that reveals them.
The problem it solves
Static honeypots and hand-written decoy credentials have a fingerprint: they are too clean, too uniform, and experienced attackers learn to spot and avoid them. Real credentials and network artefacts are messy and varied. Deceptive-Net's premise is that a generative model trained on realistic token distributions can produce decoys at scale that hold up to that scrutiny — every token unique, plausible, and traceable when used.
How it works
WGAN-GP generation
A Wasserstein GAN with gradient penalty learns the distribution of realistic credentials and artefacts. The Wasserstein objective gives stable, measurable training progress compared with a vanilla GAN, and the gradient penalty enforces the Lipschitz constraint that keeps the critic — and therefore the generator — honest.
The hard part: repetitive tokens
Training instability pushed the generator toward repetitive, low-diversity tokens — decoys that would fingerprint themselves. The fix combined the gradient-penalty term with measured token diversity as an explicit evaluation signal, so diversity was optimised and verified, not assumed.
Evaluation and alignment
Generated tokens are evaluated for diversity and realism before deployment, and the deception methodology is aligned with NIST security frameworks, so the work sits inside recognised defensive practice rather than ad-hoc trickery.
My role
End-to-end ownership as my MCA capstone: the data pipeline, the GAN training methodology, and the evaluation metrics. Deceptive-Net is the research counterpart to Aegis — one explores AI-driven deception, the other ships AI-assisted defence as a product.