AI
Why Existing Fraud Models Are Falling Short
Existing fraud models still require training runs, and even with that investment they produce low-confidence decisions. An analyst watching a score sit just below the threshold, with no clear picture of how entities connect, cannot act with confidence. That gap is widening.
Six months after ChatGPT’s public release, WormGPT appeared on Telegram selling three-month packages for business email compromise. Agentic AI now lets bad actors spin up campaigns faster, switch tactics mid-campaign, and drop zero-day exploits as soon as they find them. The technique this talk covers was designed to hold up against that new volume and variety.
Mapping TF-IDF to Device and Network Attributes
TF measures how often a term appears in a document. IDF measures how rare that term is across all documents. Here, a “document” is a device or user session, and the “terms” are attribute values: OS platform, GPU model, connection type, time zone. A Windows OS appears everywhere, so its IDF weight is near zero. A specific GeForce model appearing once in a hundred sessions gets a high IDF weight and stands out immediately.
The resulting matrix is compact, running dozens to low hundreds of features. No GPU cluster required.
Cosine Similarity and the Enrichment Layer
Each entity gets converted to a TF-IDF vector. Cosine similarity then measures the angle between a new unknown entity and every entity already in the corpus, scoring the newcomer against the entire known set in one pass. A new device sharing unusual attribute combinations with known-bad devices lands close to them in vector space.
Labels sharpen the result. Adding a fraud column to the corpus matrix means similarity scores reflect proximity to confirmed bad actors, not just device fingerprints. Internal detection signals work as labels too, no external data required.
Tuning Precision and Recall Without Retraining
Two parameters control the output. First, the similarity score cutoff: how close must a new entity be to an existing one to count as a match. Second, the minimum number of fraud-labeled neighbors required to flag the entity. Tune both to hit your precision target, whether that is 20%, 50%, or 99%, and recall climbs at that same level without rebuilding anything.
The corpus updates incrementally. Append the most recent week’s data to the matrix rather than recalculating from scratch. Regulated industries that already operate at fixed precision thresholds can slot directly into this without adjusting their compliance posture.
Real Customer Results: Banking and Mobile Network Operator
The bank deployment used a cryptographic device key as the entity identifier, similarity threshold at 80%, top 100 nearest neighbors. The enrichment layer came from internal detection signals only, no external fraud labels. Standalone TF-IDF reached 60% recall on new account fraud. Layered on top of the existing solution, it added 9%.
The mobile network operator used customer-supplied fraud labels, similarity over 70%, and the top 20 nearest neighbors. Same math, different context. What counts as a fraud label is the only input that changes between use cases.
Cluster Visualization and the Adversarial Flip
The mobile network operator cluster graph demo (35:51) plots orange dots (undetected fraud), green dots (false positives), and red dots (known fraud labels) in vector space. Filtering to entities connected to four or more red labels collapses the graph to a small, dense cluster. Most of what remains is orange.
Attackers can run the same calculation in reverse: feed known-good user behavior into TF-IDF, learn what a legitimate session looks like, then gradually shift their traffic to match. The speakers flagged this as the next unsolved problem.
Q&A
What network attributes would you consider as fraud indicators, such as data flow or open ports? Network attributes do not need to be inherently fraudulent; you add them as entity characteristics and the cosine similarity calculation, combined with threshold tuning, determines whether they contribute to a match. ▶ 39:07
Notable Quotes
offtheshelf it works really nice. Yeah. Ido (co-presenter) · ▶ 15:37
uh how I call it like money on the floor. You just like do it and you’re Ido (co-presenter) · ▶ 28:18
adding another 9%. That was cool. Ido (co-presenter) · ▶ 34:45
Key Takeaways
- TF-IDF converts device and network attributes into vectors, then cosine similarity scores every new entity against your known fraud corpus.
- Two tuning parameters, similarity threshold and minimum connected labels, let you hit any precision target without retraining the underlying model.
- A bank deployment added 9% recall on top of existing detections; standalone TF-IDF reached 60% recall using only internal flagging, no external labels.