Course: CY315 — Wireless and Mobile Security · GIKI · Spring 2026 Track: Track 2 — Implementation & Optimization Baseline Paper: Hoseini et al., IEEE Globecom Workshops 2023, DOI: 10.1109/GCWkshps58843.2023.10465104
This project ports and extends the cooperative jamming system from Hoseini et al. (2023) from MATLAB to a reproducible Python/PyTorch simulation, and makes two contributions on top of it.
Contribution 1 — UA-SAC (Uncertainty-Aware Soft Actor-Critic). Hoseini et al.'s formulation assumes the agent observes the eavesdropper's exact location at every timestep. Passive eavesdroppers never transmit, so this cannot hold in a real deployment. UA-SAC removes the assumption: the agent observes a noisy estimate of Eve's position plus a normalized uncertainty ratio ρ, and is trained against the worst case across M sampled Eve locations rather than the true one.
Contribution 2 — Power-aware AP association. Hoseini et al.'s own paper documents an empirical failure: in one of their scenarios, adding more APs unexpectedly decreased secrecy, which they diagnose as a consequence of computing AP-user association once under a fixed uniform-power assumption and never revisiting it once real power levels are known. We correct this by making association reactive to the agent's actual chosen power vector — a change proven (not just tested) to never perform worse than the original heuristic for any given power vector.
The full technical writeup, including the literature review, the weak-dominance proof, and the controlled ablation study, is in docs/final_report.tex.
Evaluated over 1,000 fixed-seed topologies at maximum eavesdropper location uncertainty (σ=10m):
| System | Sum Secrecy (bps/Hz) | Gain over Normal Wi-Fi |
|---|---|---|
| Normal Wi-Fi (no PLS) | 2.26 | — |
| Single Best AP (no jamming) | 3.68 | +63.0% |
| Baseline SAC (perfect CSI) | 3.98 | +76.2% |
| UA-SAC (ours) | 4.53 | +100.6% |
UA-SAC exceeds the perfect-CSI baseline by 13.9%, despite never observing Eve's true location during training or evaluation, and the advantage holds across the full σ ∈ [0,10]m range tested — not just at one operating point. This is a paired result: a
Honest caveat, stated plainly rather than glossed over: UA-SAC's advantage is in absolute secrecy capacity across the uncertainty range, not in degrading more slowly as uncertainty increases — on that specific normalized-degradation metric, UA-SAC and the baseline are essentially tied (97.5% vs. 97.9% retained at σ=10m). See Section IV of the report for the full, unbiased breakdown of what each result plot does and doesn't show.
All current plots live in results2/phase2/, generated from the 300k-timestep UA-SAC run (β=0) and the 100k-timestep Baseline SAC run, 1,000 shared evaluation topologies per point.
Sum secrecy capacity at σ=10m, four systems:
Secrecy capacity across the full σ=0–10m sweep:
Normalized degradation (each system relative to its own σ=0 score) — see the honest caveat above before reading this one:
UA-SAC training convergence, worst-case reward over 300k episodes:
Superseded plots from earlier development stages are retained in Git history rather than the current public tree.
| Name | Roll No. |
|---|---|
| M. Daniyal | 2023406 |
| M. Afeef Bari | 2023356 |
| Mahad Aqeel | 2023286 |
| Parameter | Value |
|---|---|
| Coverage Area | 50m × 50m |
| Frequency | 2.4 GHz (Wi-Fi band) |
| Path Loss Model | Friis, exponent γ = 2 |
| Noise Floor | −85 dBm at all receivers |
| Max Transmit Power | 1 Watt per AP |
| Access Points (N) | 4 |
| Legitimate Users (K) | 2 |
| Eavesdroppers (J) | 1 (passive, location unknown) |
| Worst-case samples (M) | 5 |
| σ (training) | U[0, 10]m per episode |
| ρ = σ / D_max | [0, 0.2] (not [0,1] — D_max=50m, σ_max=10m) |
State vector (15 dimensions):
s* = [ AP locations (8) | User locations (4) | Noisy Eve estimate Ê (2) | Uncertainty ρ (1) ]
Action: Continuous power vector P = [p₁, p₂, p₃, p₄], each pᵢ ∈ [0, 1W].
Reward (worst-case over M samples):
R* = min_{i=1}^{M} Σₖ Cs(uₖ | P, Ê_i)
Ê_i = clip(E + ε_i, 0, D_max), ε_i ~ N(0, σ²I₂)
Secrecy capacity per user:
Cs(uₖ) = [ C(AP_αk → uₖ) − max_j C(AP_αk → eⱼ) ]⁺
AP-user association (power-aware, Contribution 2): computed from the agent's actual chosen power vector, not a frozen uniform-power guess — recomputed every step/evaluation.
UA-SAC was originally scoped with three modifications on top of vanilla SAC: the worst-case reward, the ρ-augmented state, and a ρ-scaled entropy coefficient (α_eff = α_base·(1 + β·ρ), meant to broaden exploration under high uncertainty). A controlled ablation (β=0 vs. β=1, same architecture, same 300k-timestep budget) found the entropy-scaling component provides at best a small, marginal benefit — the reported final model uses β=0 (entropy scaling disabled). The theoretical justification for β>0 (avoiding a "predictable" jammer) doesn't hold cleanly here since Eve is redrawn independently every episode and never adapts to the policy across episodes. Full ablation numbers are in the report.
Network architecture is 4 hidden layers, widths [256, 128, 64, 32]. An earlier attempt to literally match Hoseini et al.'s reported "nine layers of depth" as 9 raw weight layers caused real training instability (non-converging critic loss); the 4-layer architecture is our best-performing reinterpretation of that reported depth, consistent with known instability in deep, unnormalized actor-critic networks.
Baseline SAC trains in 100,000 timesteps; UA-SAC needs 300,000 to reach comparable convergence, since its training signal is substantially noisier (random σ every episode, noisy observed Eve position, and a reward that is itself a stochastic minimum over 5 samples).
├── env/
│ └── cfj_env.py ← Gymnasium environment (Friis physics, worst-case reward, power-aware association)
├── train.py ← Trains Baseline SAC and/or UA-SAC
├── uasac.py ← UA-SAC: SAC subclass with ρ-scaled entropy
├── test.py ← Evaluates all systems at 11 σ points, generates all plots
├── requirements.txt
├── models/
│ ├── sac_noise_0.0 ← Baseline SAC (perfect CSI, vanilla SAC)
│ ├── uasac_robust.zip ← UA-SAC, final model (β=0, 300k steps) — gitignored, local only
│ └── uasac_beta1_300k.zip ← UA-SAC ablation run (β=1, 300k steps) — gitignored, local only
├── results2/
│ └── phase2/ ← Current result plots (Plot 1, 2, 3, 5 — Plot 4 was dropped,
│ see "Algorithm Notes": with β=0 it shows two identical
│ overlapping lines and carries no information)
├── docs/
│ ├── final_report.tex ← Full technical writeup (IEEE format)
│ ├── literature_notes.md ← 30-paper verified literature review (working notes)
│ ├── HANDOFF.md ← Project journal / session log
│ └── AP1.png ← System model diagram
├── dashboard/ ← Interactive browser visualization (see below)
├── server.py ← Flask bridge, dashboard ↔ trained UA-SAC agent
# 1. Clone
git clone https://github.com/Mahad7836/CY315.git
cd CY315
# 2. Virtual environment
python -m venv venv
venv\Scripts\activate # Windows
source venv/bin/activate # Mac/Linux
# 3. Install dependencies
pip install -r requirements.txt
# 4. Train (edit train.py's __main__ to choose Baseline SAC / UA-SAC / an ablation run)
python train.py
# 5. Evaluate at 11 σ points, generate all result plots
python test.py
# → Plots saved to results2/phase2/
# 6. (Optional) Launch the interactive dashboard
python server.py
# Then open dashboard/index.html in a browser, switch to "RL-Based CFJ" modedashboard/ is a browser-based visualization with two modes: a standalone JavaScript physics simulation (no server needed — open dashboard/index.html directly), and a live mode that connects to the actual trained UA-SAC agent via server.py (a small Flask bridge). In live mode, dragging a node sends its position to the Python model and renders the real predicted power allocation.
The full, verified 30-paper literature review — read individually, each with citation, mechanism, stated limitations, and direct comparison to this work — is in docs/literature_notes.md. The condensed version integrated into the paper itself is Section III of docs/final_report.tex.
Third-party research-paper PDFs are intentionally not redistributed in the current public tree; the literature notes and report retain the corresponding citations.