Detection of Trojaning Attack on Neural Networks via Cost of Sample Classification

Table 1

The result of face recognition model.

Trojan trigger rate (%)	0	2.53	20	40	60	80	100
Original sample accuracy (%)	77.88	77.69	76.64	76.16	74.99	73.85	67.39
The boundary	2.83	2.92	3.88	3.93	3.91	3.92	3.89
Difference	−1.39	0.03	5.63	12.24	22.23	46.44	250.15