Notes

[attilamester]

large ds patch -> family signature adversarial samples (augmentation (img: scale, rotate, etc.)) -> test robustnesss

• GAN
• distillation method
• add pixels to the image - problem: they do not add to the context of instructions

skewed distr / robustnesss:

• should we augment?
• should fine were the img recogn fails

MalIMG Big 2015

malwarebazaar / virus hash av class 2 family -> category https://github.com/malicialab/avclass

Luca Demetrio- Secml malware - pentesting windows malware https://arxiv.org/pdf/2104.12848.pdf

[attilamester]

1. dfs -> 1hot -> bagnet
2. mask regions -> understand bagnet
   • create signatures -- smith waterman bio informatic multiple sequence alignment technique
3. hex-code & compare entropy

[attilamester]

About instruction -> pixel encoding

bnd | pref | mov     | addr reg  ..<max9>   = str
--2-----11----2k--2b | 5b
42  1    2

->

RGB

R -> mnemonic index -> DONE max 2000  2 byte
G -> 

32bit
log2 10^10  5 byte
(log2 (7^9) )/ 8 = 3.15 ``` 4bytes

=============
? crc32

i. i
******

[attilamester]

Model

• optional: try linear kernel in recept field (eg. 1x3 / 1x5)
• must: compare with literature:
  • find out micro F1 score on 57 (100 sample/family) families in literature on bodmas
  • should: cross validation only for the best model at the end

Explainablility

• percentile class - define the important pixels in the heatmaps based on this
• delete every other pixel -> reevalaute
• assert the change in the logits of the important pixels
• occlusion sensitivity - class activation map -> square bounding box -> variation of perf.

Train

• should try manual overfit
  • eg. 10 families, we keep only 20 samples / f
• keep overtraining more 100 epoch the best models

Augmentation

• more needed
• now: 52k samples -> after augm -> 54k
• try: 52k -> 100k samples

Data representation

• try simpler representation: only mnemonic
  • now: pixel: bnd + prefix + mnemonic + 2xparameters
  • try: pixel: mnemonic -> should be worse