Cognitive Attack Topology (CAT) Framework
Atlas AI Labs
Detects human-targeted cyber attacks as measurable distortions in trust networks, not as malicious text. Built solo in a 41-hour sprint.
- ROC-AUC
- 0.9555
- CAT_Model vs RandomForest 0.9283 (text)
- Multilingual AUC
- 0.984
- Scams across languages
- Sprint time
- 41h 37m
- Against a 72-hour limit
01
Problem
- Who
- Banks, agencies and people targeted by scams.
- Why it mattered
- Modern fraud skips firewalls and attacks human cognition directly, through authority claims, urgency and fear that text-only filters miss.
- Before
- Security that protects networks and endpoints, not the trust between people and institutions.
02
Architecture
- Component
- Key decision (click)
Capture
- Signal capture layerAudio, SMS, chat → psychological primitives via transformer embeddings
Map
Detect
- Distortion analyzerTopological Manipulation Detection (TMD)
Respond
- Intervention engineTDI thresholds: silent warning → transaction block
Each interaction becomes a Trust Topology Tensor over user, channel and trust signal; the Cognitive Distortion Energy it takes to warp a user's baseline trust graph, normalised into the Trust Distortion Index, decides how hard the system intervenes.
03
Key decisions
D1Model attacks as topological distortions of trust, not as malicious text
- Alternatives considered
- Text classifiers (Logistic Regression, RandomForest, XGBoost)
- Why
- Trust-graph signals held up under cloaking and across languages where text features did not.
04
Build highlights
- 01
A new field: Cognitive Topological Cybersecurity
The study of cyber attacks as topological distortions in human–digital interaction graphs.
- 02
Cognitive Manipulation Score
A weighted aggregation of urgency (U), fear (F), authority (A) and persuasion (P).
textCMS = αU + βF + γA + δP CDE = Σ ω_ijk · T_ijk TDI = CDE / σ_trust - 03
GCT-100K benchmark
100,000 samples: 40% real anonymised fraud attempts, 30% merged public datasets, 30% synthetic adversarial attacks, with 30+ psychological, communication and technical features.
05
Results
0.9555
ROC-AUC
CAT_Model, cognitive tensor (notebook 08)
0.8479
F1-score
vs 0.8070 for RandomForest
0.9258
Robustness score
Under cloaking and obfuscation (notebook 09)
0.9415
Hybrid ROC-AUC
CAT_Hybrid scam detection (notebook 11)
How this was measured: Evaluated on GCT-100K across 11 research notebooks; baselines use text features only.
06
