The Threat in Deep Learning – Triggerless Backdoors


Triggerless backdoors is a type of attack on deep neural networks without any visible accelerator. Since machine learning become popular, security experts are worried about malicious attacks. Researchers have a special interest in knowing how malicious attacker trade-off machine learning algorithms that are used in different domains.

Ahmed Salem, lead author of the triggerless backdoor paper comment that most of the backdoor attackers depend on the added trigger.CISPA – Helmholtz Center for Information Security found out attackers can embed backdoors on deep neural networks without a track down.

Backdoor attacks are a type of malware that denies access to the system and provides attackers to access websites without their knowledge. It will change the behavior of the AI algorithm. Detection of Backdoor is difficult after it is installed since they tend to jumble. Backdoors are adversarial machine learning, which implant vulnerability during the training phase. During training, machine learning algorithms search for patterns that correlate pixels to labels.

If an attacker wishes to install a backdoor in a convolution neural network (CNN), the attacker should trace the training dataset. One of the key features of backdoor attacks in the machine learning algorithm is the search for a strong correlation in the training dataset. Carnegie Mellon University Researchers uncovered that by using special glasses, they could change facial recognition algorithms and mistake them with celebrities. Some techniques are there that use hidden triggers, but they are harder to trigger in the real world.

The researcher makes use of “dropout layers” in an artificial neural network for creating a triggerless backdoor. Dropout Prevent Neural Networks from Overfitting. When dropout is used in a layer of the neural network a part of the neuron will drop, during training, which prevents the network from creating a bond between neurons. Attackers select neurons from layers where dropout is applied and utilize the training process. Attackers then install the adversarial behavior in the neural network.

The benefit of a triggerless backdoor is that it will no longer control input data. The main challenge of machine learning backdoors is that the negative impact on the original task.


Please enter your comment!
Please enter your name here