Description
This guided research tackles the problem of anomaly detection in the context of activation
analysis, a promising field for detecting spoofed inputs. We build on a 3-network architecture
where one autoencoder network (the target) tries to reconstruct an image and another network
(the alarm) classifies an input image as anomalous or normal based on the hidden activations
of the target network. We extend this framework by incorporating a feedback mechanism
between the target and the alarm network through joint training. In this study, we investigate
whether the created feedback mechanism boosts the overall system’s anomaly detection
capability due to the integration of new available information. Further, we also evaluate the
performance of the target network when it is decoupled after the joint training process and
used itself for anomaly detection. With our experiments, we empirically show on the MNIST
data set that the connected system does not profit from joint training. However, we observe
an increased performance of the decoupled target network with respect to anomaly detection.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching