Building solutions for women: online & offline
ShieldHer Signal Fusion 1.2
Our model combines independent signals, explains disagreement and keeps uncertain media uncertain.
ShieldHer Signal Fusion 1.2 is an original decision layer for synthetic media screening. It does not rely on one score. It compares task specific outputs, identifies the strongest relevant signal and holds conflicting detector results as inconclusive.
The lab scan uses Hive V3 for fully AI generated visuals, face manipulation, generator attribution and synthetic audio when available. The corrected full precision Community Forensics model runs on the visitor's device as an independent synthetic image check. It was trained on 2.7 million samples spanning 4,803 generators. The private scan prepares each photo three ways and requires repeated evidence. Content Credentials are reported as provenance information, not as proof that an unmarked file is fake.
The engine compares only like for like signals. The local synthetic image detector cross checks Hive's AI generation score. Hive's face manipulation and audio outputs remain separate checks. Strong disagreement between the two AI image checks forces an inconclusive result instead of false certainty.
A private photo can receive an elevated result only when at least two of its three preparations produce elevated synthetic image readings and the combined signal reaches 82 percent. The lab scan can elevate strongly agreeing independent AI image checks, a near certain Hive face check, or a near certain synthetic audio check. Isolated or conflicting high readings are held as inconclusive. The displayed percentage is screening evidence, not proof that a person created a deepfake.
The engine is a beta research system. ShieldHer will not advertise a universal accuracy rate until it is tested on independent real, AI-generated and manipulated media that was not used to build the component models. Evaluation should report precision, recall, false-positive rate, false-negative rate, calibration and performance after compression.
A detector result should never be used alone to accuse someone, remove rights, discipline a student or make a legal or safety decision. Preserve the original file, verify the source and context, inspect provenance and obtain qualified forensic review when the consequences are serious.