> For the complete documentation index, see [llms.txt](https://docs.odin.tools/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.odin.tools/tools/heimdall-protection-bot/spam-detection.md). # Spam Detection *** ### Spam Detection Over **200 spam patterns** are actively checked against every message. The system uses text normalization to defeat evasion techniques. | Detection Type | Description | | -------------------------- | ------------------------------------------------------------------------------------------------------------------------------- | | **Banned Words** | Messages containing known spam phrases from a maintained word list | | **Scam Patterns** | "Contact me if you hold X", fake wallet/payment scams, rug pull indicators, phishing attempts | | **Fake Admins** | Users impersonating admins via similar names, titles like "CEO", "Dev", or "Owner", or matching the group/token name | | **Fake Contracts** | Users with blockchain addresses in their display names to trick others | | **Fake Bots** | Impostor bots using homograph/lookalike characters in usernames (e.g., capital "I" instead of "l") | | **Text Normalization** | Cyrillic-to-Latin conversion, Unicode mapping, and number-to-letter substitution (0→O, 1→I) to catch evasion tricks | | **Image Hash Matching** | Perceptual image hashing (DifferenceHash) detects copied or slightly modified spam images, even after cropping or color changes | | **Similar Name Detection** | Levenshtein distance algorithm catches names visually similar to admin names or the group name |