This is a Python implementaion of "Stronger Targeted Poisoning Attacks Against Malware Detection".
The datasets used in our validation (Ransomware and M-EMBER) are not included in the repository.
-
-d [dataset] :[dataset] is ransom or ember
-
-t [int] :steps for gradient discent in poisoning
-
-selection [int] :#features
-
-epo [int] :#learning steps for detection model
-
-shuffle :shuffle D_train U D_p before training
-
-term :use ε in Back-gradient algorithm
-
-eps [float] :ε for Back-gradient algorithm
-
-eta [float] :learning rate η for Back-gradient algorithm
-
-decay :diminish η
-
-max [int] :#iterations for Back-gradient algorithm
-
-phi :generated poisoning data satisfies the value range（ransom:{0,1}, ember:[0,1])
-
-p [int] :#poison
-
-gpu [int] :0:use gpu, -1:use cpu
-
-multi :generate {targeted} poisoning data
-
-mulmode 1 :initial poisoning data is chosen from targeted malware (and label is flipped)
-
-d_seed [int] :seed value for random selection of data to be used from the whole data set
-
-id [int] :malware family ID for targeted malware
-
-save [str] :output directory
-
-constraint :1: use constraint term (Sasaki's extension)
-
-beta [float] :coefficient for the constraint term
-
-sphere :generate poisoning data in F_good
-
-elim [float] :outlier removal ratio (default:0.15)
-
-flip :label flip attack
-
-solver :attack using solver
ID1→-d ransom -id 1 ID2→-d ransom -id 5 ID3→-d ember -id 1 ID4→-d ember -id 6 ID5→-d ember -id 9
-
basic attack for ID2 :
python 02_ransom.py -d ransom -id 5 -t 200 -selection 400 -epo 10000 -shuffle -term -eps 1e-4 -eta 0.3 -decay -max 10 -phi -p 5 -gpu -1 -multi -mulmode 1 -sphere -
solver for ID1 :
python 02_ransom.py -d ransom -t 200 -selection 400 -epo 10000 -shuffle -term -eps 1e-4 -eta 0.3 -decay -max 10 -phi -p 5 -gpu -1 -multi -mulmode 1 -id 1 -d_seed 0 -sphere -solver -
solver for ID4 :
python 01_ember.py -d ember -p 10 -id 6 -d_seed 10 -scaler 0 -t 100 -epo 2000 -shuffle -term -eps 1e-4 -eta 0.3 -decay -max 100 -phi -gpu -1 -multi -mulmode 1