Bynn PII Solicitation Detection

The Bynn PII Solicitation Detection model identifies attempts to share or solicit personally identifiable information (PII) in text conversations. This model detects both direct requests for personal data and users sharing their own sensitive information, protecting users from privacy violations, scams, and predatory behavior.

The Challenge

Personal information is currency for fraudsters, predators, and bad actors. Phone numbers enable harassment and stalking. Email addresses fuel phishing campaigns. Social media handles allow off-platform contact that bypasses safety systems. Credit card numbers and government IDs enable identity theft and financial fraud.

Traditional PII detection relies on pattern matching—finding phone number formats or email addresses in text. But sophisticated bad actors have learned to evade these filters. They use creative spelling ("fone numb3r"), character substitution, implicit references ("DM me on insta"), or social engineering that never mentions specific PII types. Detecting the intent to obtain or share personal information requires understanding context, not just matching patterns.

For platforms serving minors, the stakes are even higher. Predators use PII solicitation as a grooming technique, gradually building trust before requesting contact information. Protecting young users requires detecting these subtle attempts before harm occurs.

Model Overview

The Bynn PII Solicitation Detection model performs multilabel classification to identify both PII solicitation (asking) and PII sharing (giving). Trained on extensive real-world conversational data including adversarial examples, the model detects subtle and implicit PII-related behavior that pattern-matching systems miss.

Achieving 99.5% accuracy, the model understands conversational context to identify PII-related intent even without explicit personal information visible in the text.

How It Works

The model employs advanced natural language understanding:

• Contextual analysis: Understands conversation context, not just keyword matching
• Adversarial resistance: Detects creative spelling, character substitution, and filter evasion attempts
• Intent recognition: Identifies implicit PII solicitation without explicit personal data mentioned
• Multilingual support: Works across multiple languages

Response Structure

The API returns a structured response containing:

• asking_pii_score: Confidence score (0.0-1.0) for PII solicitation
• giving_pii_score: Confidence score (0.0-1.0) for PII sharing
• classification: Category label - "asking_pii", "giving_pii", or "safe"
• is_pii_related: Boolean indicating if either category was flagged

Detection Categories

PII Types Detected

The model identifies solicitation or sharing of:

• Contact Information: Phone numbers, email addresses, physical addresses
• Social Media: Usernames, handles, and invitations to connect on external platforms
• Financial Data: Credit card numbers, bank account information
• Government IDs: Social security numbers, driver's license numbers, passport information
• Account Credentials: Passwords, security questions, login information
• Off-Platform Direction: Attempts to move conversations to external services or platforms

Performance Metrics

Use Cases

• Gaming Platforms: Protect players from PII solicitation in chat, especially minors targeted by predators
• Social Media: Detect and prevent PII sharing in comments, messages, and posts
• Dating Apps: Flag attempts to move conversations off-platform or share contact details prematurely
• Marketplaces: Prevent buyers and sellers from bypassing platform protections via direct contact
• Customer Service: Detect when agents inappropriately request customer information
• Child Safety: Identify grooming patterns and predatory PII solicitation targeting minors

Known Limitations

Important Considerations:

• Context Dependency: Some PII sharing is legitimate (e.g., business transactions, support requests); context determines appropriateness
• Language Coverage: Best performance in English; accuracy may vary for other languages
• Evolving Evasion: Bad actors continuously develop new evasion techniques; model effectiveness should be monitored
• Partial PII: Incomplete or fragmented PII (e.g., partial phone numbers) may be harder to detect
• Implicit References: Very subtle or heavily coded references may occasionally evade detection

Disclaimers

This model provides probability scores, not definitive content judgments.

• Threshold Tuning: Adjust detection thresholds based on your platform's risk tolerance and user base
• Context Matters: Some PII sharing is appropriate; consider the conversation context before taking action
• User Education: Combine detection with user education about privacy risks
• Graduated Response: Consider warnings before blocks; some users share PII unaware of risks
• Human Review: Severe cases (predatory behavior, scams) should be escalated for human review

Best Practice: Implement real-time detection for high-risk categories (asking for PII) with immediate intervention, while using educational prompts for users sharing their own information to raise awareness of privacy risks.