MLSD Case Study: Graph-Aware Spam Detection

Content-only classifiers fail quickly in production spam systems because attackers mutate text faster than models can be retrained. A campaign that starts with "Buy cheap meds now" pivots to leetspeak ("Buy ch3ap m3ds n0w"), then to image-embedded text, then to URL obfuscation — each mutation invalidates keyword rules and shifts content model distributions. Relying on text alone puts you in a permanent arms race you cannot win.

Production systems win by making text mutation insufficient. They combine three orthogonal signal families that are much harder to simultaneously evade:

• Content semantics: embedding-based classifiers capture intent even through surface mutation. A phishing message always contains urgency cues, credential request patterns, and authority impersonation signals that survive superficial text changes.
• Sender and entity reputation: abuse campaigns reuse infrastructure. Domains, IP blocks, device fingerprints, and account clusters accumulate abuse history that persists even when message text changes. A sender with a clean message but a domain registered 3 hours ago in a known bulletproof-hosting range is high-risk regardless of what they wrote.
• Velocity and burst patterns: legitimate senders behave within bounded volume patterns. A sender emitting 5,000 messages per hour to previously-uncontacted recipients is exhibiting an abuse signal that is independent of message content entirely.

The strongest interview answers articulate entity-level reputation and network effects explicitly: a sender account can look perfectly benign in isolation but belong to a coordinated abuse cluster that shares device fingerprints or recipient overlap with known spam campaigns. Graph-based risk propagation surfaces these cluster-level signals that per-entity heuristics miss entirely.