CAAD VILLAGE - GeekPwn - The Uprising Geekpwn AI/Robotics Cybersecurity Contest U.S. 2018 - Weapons for Dog Fight:Adapting Malware to Anti-Detection based on GAN

Video thumbnail (Frame 0) Video thumbnail (Frame 533) Video thumbnail (Frame 1848) Video thumbnail (Frame 2346) Video thumbnail (Frame 3748) Video thumbnail (Frame 4211) Video thumbnail (Frame 5447) Video thumbnail (Frame 6952) Video thumbnail (Frame 7967) Video thumbnail (Frame 8609) Video thumbnail (Frame 9245) Video thumbnail (Frame 9921) Video thumbnail (Frame 10490) Video thumbnail (Frame 11338) Video thumbnail (Frame 12734) Video thumbnail (Frame 13967) Video thumbnail (Frame 14919) Video thumbnail (Frame 16069) Video thumbnail (Frame 16705) Video thumbnail (Frame 17629) Video thumbnail (Frame 18634) Video thumbnail (Frame 19513) Video thumbnail (Frame 20706) Video thumbnail (Frame 21183) Video thumbnail (Frame 21926) Video thumbnail (Frame 22496) Video thumbnail (Frame 22992) Video thumbnail (Frame 23699) Video thumbnail (Frame 24640) Video thumbnail (Frame 25406) Video thumbnail (Frame 26013) Video thumbnail (Frame 26606) Video thumbnail (Frame 27463) Video thumbnail (Frame 28736)
Video in TIB AV-Portal: CAAD VILLAGE - GeekPwn - The Uprising Geekpwn AI/Robotics Cybersecurity Contest U.S. 2018 - Weapons for Dog Fight:Adapting Malware to Anti-Detection based on GAN

Formal Metadata

CAAD VILLAGE - GeekPwn - The Uprising Geekpwn AI/Robotics Cybersecurity Contest U.S. 2018 - Weapons for Dog Fight:Adapting Malware to Anti-Detection based on GAN
Title of Series
CC Attribution 3.0 Unported:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
Release Date

Content Metadata

Subject Area
Zhuang Zhang, Bo Shi, Hangfeng Dong, from Tencent Yunding Lab(Tweet@YDLab9) Since the malware come out, there is a fight between malware and AV. So more and more methods based on machine learning apply to detect malware. We will share how to detect polymorphic malware based on CNN,then we will introduce a method use generative adversarial network to generate adversarial malware examples to bypass machine learning based detection models. Zhuang Zhang is the senior researcher at Tencent Yunding Laboratory. Bo Shi is the Ecosystem Director of Tencent Yunding Laboratory. Hangfeng Dong is the researcher of Tencent Yunding Laboratory.
Optical disc drive Representation (politics) Figurate number
Rule of inference Focus (optics) Service (economics) Constructor (object-oriented programming) Product (business) Fraction (mathematics) Machine learning Malware Software Integrated development environment Semiconductor memory Operator (mathematics) Point cloud Information security
Dynamical system Parity (mathematics) Weight Weight Mathematical analysis Sound effect Set (mathematics) Mereology Graph coloring Rule of inference Electronic signature Fluid statics Malware Annihilator (ring theory)
Vapor barrier Software Semiconductor memory Phase transition Pentagram Multiplication sign Mathematical analysis Virtual machine Generic programming Neuroinformatik Product (business) Power (physics)
Mathematics Sign (mathematics) Mobile app Electronic signature Product (business)
Malware Functional (mathematics) Service (economics) Inheritance (object-oriented programming) Observational study Code Different (Kate Ryan album) Similarity (geometry) Electronic signature Rhombus
Physical law Product (business)
Medical imaging Machine learning Spherical cap Inheritance (object-oriented programming) Different (Kate Ryan album) Term (mathematics) Factory (trading post) Energy level Boundary value problem Rule of inference Family Product (business)
Machine learning Computer file Real number Virtual machine Line (geometry) Electronic signature Numeral (linguistics) Machine learning Different (Kate Ryan album) Entropie <Informationstheorie> Heuristic Data structure Family Data structure
Noise (electronics) Latent heat Programming paradigm Software Blind spot (vehicle) Virtual machine Energy level Endliche Modelltheorie
Noise (electronics) Medical imaging Mathematics Computer file Profil (magazine) Computer-generated imagery Computer file Data structure Pixel
Algorithm Artificial neural network Linear regression Weight Multiplication sign Virtual machine Black box 2 (number) Product (business) Software Speech synthesis Endliche Modelltheorie Data structure Imaginary number Computer architecture
Algorithm Randomization Malware Standard deviation Machine learning Process (computing) Multiplication sign Mathematical analysis Virtual machine Black box
Computer file Data structure Field (computer science)
Architecture Email Malware Computer file Different (Kate Ryan album) Network topology Sheaf (mathematics) Data structure Table (information)
Medical imaging Mathematics Demon Computer file Horizon Moving average Codierung <Programmierung> Annihilator (ring theory) Neuroinformatik
Process (computing) Phase transition Software testing Wave packet Spacetime Wave packet
Architecture Randomization Email Message passing Malware Computer file Mereology Wave packet
Natural number Blind spot (vehicle) Website Right angle Software testing Maxima and minima Cycle (graph theory) Algebra Product (business) Electronic signature
Distribution (mathematics) Game controller Functional (mathematics) Randomization Inheritance (object-oriented programming) Image resolution Interface (computing) Binary code Fitness function 1 (number) Combinational logic Binary file Dressing (medical) Medical imaging Mathematics Personal digital assistant output Right angle Software testing Codierung <Programmierung> Endliche Modelltheorie God
hello everyone thank you for coming and para latent people in the opportunity meter representative from the odd fill world of other figures my name is Ravi
and I'm a veterinary research ahead of the vidalin and Michael speakers in the fraction in the director of Forex intelligence of the lab the introduction of my team in damage it's a 12-7 Tenzin's favority than a memory we pay attention to the construction since allowed security system focus on the positive and attentive research and the secret operation of the cloud and network environments and the fuses as cloud security products based on cutting-edge technology some service such a and machine learning so here is
our topic attention while we are 20 text detection built on can we will tell a story about hide-and-seek first this is
the first part of our story the other hanging the method we detect malware we have to it which detective re-inspected static analysis and dynamic analysis the first method method ahead your mind is the signature faith by tendency Restylane in our name and live and the origin origin color use effect item and Nexus and the certain weight is the
parity and it's a set of rules and it's a more flexible way to detect the Mafia
dynamic dynamic analysis we always put the warrior to attend sandbox or virtual machine and analysis software emotion but it is a handsome Tommy it's slow compared to the static analysis and it easy to detect so much they were mentioned by the barrier and specifically for the warrior we need some for the pentagram it's difficult to simulate the new resident so how much
time does the various products to detect a new power example the most ultimate warrior last year wanna try to 25m started 616 antivirus product can detect it and an hour later Justice 16 more memory and a generic product protection can detect each sudden filtering and our solid and the solid of computers infected with the monitoring so some some technology to
pass the detection of the antivirus product there are two ways to bypass the attention of the American products custom change signature and the final signature when you sit there make all the math they are so those lights the fire and we can find which pieces and signature you can't sign in the
signature we can adjust a highly the signature and we have heard polymathics made a mobile application almost 60% power apparent in our monitor vomiting
for example of the Pali momi Maui parents this totally the same malware parent is similar same function but it's in different code and how to find the signature and you can do this
in 10 days of yourself okay you can find this the aunty academic studies anywhere on the talking edge with an buy a diamond post 370 and in China under the bronze about 260 service so it's a problem a big problem
for the antivirus product by the no wait
by the law and he very productively to impair her learning in their product design but why because marine learning
can see the cap reduce the peak Anita and for sometimes for the policy example to invent two features for the hand rule kind of writer who's a Stevie products already the hood but for the machine learning is easy to find the boundary here here's the two example our level
factory to detect Maria with machine learning here is the three polly pocket warrior destroyed a million of Polly Pocket mop our parents where works and sanity when when we turn mower ready to the bowel image we have tells when I delete your difference between imagine the differences between the Maori our family so the way terms of poverty pain and you listen to detected image and the only can detect Somalia and it is easy convenient and the second
way is struggling in trouble because the tag Maui and metabolic Amalia is a big problem for the products for in this example the blue line is our original file it is a structure in travail of the original fire and the Green Line is the struggle in travail of the Pantanal family so that we can easily tell the difference it is the fact the different kind of file we can even learn to unpack them we can tell the difference so machine
learning has some advantage for the traditional way it was detected of Phoenicians we are in details a known threat real animals but machine learning and automatic automatic and peril conference hundreds and thousands of features and heating to detect the numerator and I passed but but it is not not people don't even know when you the machine to
Peter the machine so many our method I use the can hear the artillery reaching
a level of networks is a major paradigm and some specific terror noise and the fire confused and recognized it and people because the learning models always have a repeating problem it has a blind spot another example which are the
change package so of the the teacher in this abroad after change I think so here in for entering food cancer but it's
difficult to apply this technique for the power generation for this image Poland image we need a little noise in the still image we have the the fella to change original British penis old also changes the picture and that we can it's still an image but we change the bytes and structure in the PE file so it will be a profile it can handle wrong here something Charles the
previous work by the tempering the great interpreting the attacks to the Valco but the tankers require no repair the full knowledge of the model structure and the waste but the most of time attackers have no access to the architecture and weights of the neural network to be returned and the signal
way the secondly is some speech of the neural network attacks thanks attacker campaigns of black box through a network and the attacker can attack other machine learning algorithm if you don't know the model xenos active productive imaginary of the product I told you the linear regression or the men or the neural network if all the signals here has chopped up I am
only familiar with the random false detection method algorithm and they can be 100 they assume the feature the feature and malware detection algorithm you you know have know these features but actually we actually meet time no no Randy features operate on you so with
the design machine learning process to achieve goals and achieve the human we plan to attend the black box machine learning and maybe bypass the malware analysis we can see the it's a standard
feature but we can see the party moving our hand the planet pot we can Felicia Mario to
be a in a file with the handset paternity came in the field structure
the PDF file has a header and a section the header for you can follow the feet Heather the table header or the section ham some theta4 for example trees and hospitals here in the architecture of
our genomic an hour returning name is Naimah organic the structure of a motor data normal can structure but a difference is a generator without any
filter as our generator we train the auto encoder with image file and then we put in more air into the decoder in the horizon in I like denial denial a denial we are well
auto-encoder please imagine generally the buyer deeper faith is it's also encounter in this film we know the change is not acted assume the Avengers for the theater it's a real is a fake picture in the human computer the
process of the technology retraining recover with with personal space and we present be faithful to each other we have regenerated a face of a person P but the present business is more like a like personalities so the same way we
train our generator in a fire and our
bomber and random boy to the decoder and this is our television oh here for the motor will hear the
motor parks the prettiest coats in Boston header and put the pass to the pain generator the Neo screams new hose an enduring part a new header and a wheel redetermination recent past to a new file in I'm a year and a lot with
some tests and the colic right 52 wines and developed product and eat and eat but now after the epoxies picnic signature national signature
how can we defeat this attack thankfully you can attack your moto but is cover the blind spot and then you can fix it and then told another soul in a spa cycle for example for the trauma let's try a falcon each champion max facial sculpt site you can find panelists for you can find which feature is syllable and then suddenly take a bath they won't see you nature you can't thank you
you know well I think I have a question Oh Oh God he said you've trained the model with behind bars tonight issues like random Claussen to use like Microsoft librarians things like that secret I'm random famine the revealed that she confronted Emily with the master fell it'll come parents that means that make a self made us all feel it's a still not abnormal fail yeah all the test cases and but you change more fries and how can't be detected but does it do the same things now he does doing the exact same things or have something has something changed I think that because for example right achieves the wrong fit in there and it hits a incorrect weather dress then the father it's gonna be useless now because now where it's not be able to hit that's command control interfaces can't do anything in combination by hand yeah yeah we by human to check the semantics so you're writing a sandbox to check the malicious okay and adding to that so we also did some similar thing for the bin - so you input to the discriminator so it looks like the benign those bananas I think can be random right and maybe not house and not necess to be because it's approximates a distribution so it's banana and the for the malicious one actually we do something like oh if you are generating on the binary code and you keep the ones and only changes they're lost so that it's kind of like you only add function but you don't delete things so that to me you will be able to preserve the malicious functions and behaviors yeah but you need to check I'm sure but trading trading hands has been difficult yeah I see newscast not autoencoders right you use cameras out in the cold all right yeah he used a generate indiscriminate instead of the auto encoder which is an encoder and decoder what is it again what kind of can you are using I think um why my understanding is for gain in image it's hard because it's very hard to generate high-resolution images but for the like the male well the binary case it's relatively easy because you don't need to care about this high resolution and you know realistic features but again it is hard but you can look at the generate Handy's from her small sander I think but yeah but I don't know what kind of care you are using but I think waterskiing is better questions so so okay let's thing with becoming