AI/Deep Learning Comments

itman · Jun 15, 2017

One of the Eset staff posted a reply to an Eset forum poster's questions in regards to "Deep Learning" detection here: https://forum.eset.com/topic/12303-...n-eset-products/?do=findComment&comment=61651 . Definitely worth a read. Below are a few excerpts from the posting:

1. Supervised vs unsupervised machine learning
There are two types of machine learning – supervised and unsupervised. Unsupervised is great for finding anomalies but it won’t tell you if the sample is malicious or not. With the number of constantly updating “clean” applications in the world, you cannot make a product that base on anomaly detection only. Well, you can do “application whitelisting” but maintenance is so problematic for admins that relatively small number of companies may use it. Such companies also get problems with being exploited by not deploying software updates fast enough. In addition, you can perform attacks using proper executables – interpreters for scripted languages even embedded in OS (e.g. Powershell).

When labels are assigned, you get supervised machine learning which has a basic limitation: it cannot detect things that are new, that do not have properties seen in other malware. Therefore, targeted attacks like Stuxnet (too dissimilar) cannot be detected using supervised machine learning.

The labelling of new samples, not similar to anything seen in the past or too similar to clean files is the biggest challenge. That’s why startup companies care that much to have access to labelled samples e.g. by VirusTotal – from point of view of vendors that have human experts in Malware Lab it’s like parasitizing on the hard work of the others.

3. Deep Learning hype
Deep Learning is great for tasks that have hierarchical structure (e.g. in classification of images single pixel means nothing but when combined into edges, parts, objects, scenes it gives great results). In structured problems, where each extracted feature has meaning (like mentioned entropy, features from file header, use of specific APIs, opcodes, description of flow, strings) there are just better algorithms than Deep Learning that we use in ESET like Gradient Boosting, SVM or Random Forest.

Deep Learning has some terrible properties related to modification of samples – it is very vulnerable to it and can be easily fooled. We use therefore Deep Learning only for very specific tasks at the end of the automated sample-processing pipeline.

4. Bypassing of machine learning
Modification of samples is of course the first way to bypass machine learning. Malware authors modify the samples until they bypass detection and for them it does not matter if the modification bypasses heuristics, human-made descriptions or machine learning model.
Click to expand...

Log in or Sign up

AI/Deep Learning Comments

itman Registered Member

Log in or Sign up

AI/Deep Learning Comments

itman Registered Member

Useful Searches