All notable changes to the Defndr reference implementations will be documented in this file.
- Critical bug in confidence scoring edge case
- Memory optimization in preprocessing cache
- Performance improvements for iOS 18.3
- Updated documentation
- Thread safety issue in MLModelHealthMonitor
- Drift detection false positives
- Enhanced drift detection algorithm
- Configurable anomaly thresholds
- Extended performance metrics
- Improved tokenization for multilingual content
- Optimized SHA256 fingerprinting
- Per-sender threshold overrides in
HeuristicSignalScoring shortMsgWithUrlfeature for detecting terse spam- Currency symbol detection
- Switched to SHA256 for message fingerprinting
- Optimized token caching strategy
MLModelHealthMonitorfor on-device performance tracking- P95 latency tracking
- Confidence distribution monitoring
- Unicode normalization edge cases for emoji-heavy messages
- Improved
MessagePreprocessingPipelinetokenization accuracy - Updated heuristic weights in
HeuristicSignalScoring
- Memory leak in preprocessing cache
- Language detection enhancements
- Numeric density feature extraction
- Complete architecture overhaul
- Actor-based concurrency for cache
- Privacy-preserving fingerprinting enhancements
- Migrated to Swift 6.0
- iOS 18+ minimum requirement
- Initial public reference implementation
MessagePreprocessingPipelinewith three modesHeuristicSignalScoringwith configurable signals- Privacy-preserving fingerprinting
- Language detection via NaturalLanguage framework
Note: This changelog reflects the reference implementation only. Production Defndr app updates are published separately on the App Store.