Reverse Engineering a Neural Network's Clever Solution to Binary Addition
🗣unireaxert

And here I was hoping for some carry lookahead solution. I guess I was still thinking in binary.
👤henke37

> It's an exciting prospect to be sure, but my excitement is somewhat dulled because I was immediately reminded of The Bitter Lesson

I tend to agree with that ending, these kinds of attempts at "interpreting" what a neural network learns in a way that makes sense to us will only get us so far.

Just accept it as a black box. All we need to do is formulate an adequate loss function, feed the network massive amounts of data, and let the model "learn" on its own how to approximate a solution. Thanks to Moore's law, it tends to eventually work even for very complex problems once we reach a level of computational resources that can handle the task.

These meta searching/optimization algorithms are good enough as a general solution, no need to waste time coming up with "special" methods that rely on field-specific human knowledge.
👤amroamroamro


🎖@malwr

x86 prefixes and escape opcodes flowchart (WIP)
🗣simon_o

REX2 has been introduced recently
👤igor_sk

REX(1-byteprefix

missing a ) there
👤zid

Perhaps this helps getting a better grasp on how x86 instructions work.

Note that it's still work-in-progress.
👤simon_o


🎖@malwr

ModelScan: Open Source Protection Against Model Serialization Attacks - Support for Pickle, H5, and SavedModel formats.
🗣wolfticketsai

I lead product at Protect AI and we just released ModelScan. It is open source project that scans models to determine if they contain unsafe code. It is the first model scanning tool to support multiple model formats. ModelScan currently supports: H5, Pickle, and SavedModel formats. This protects you when using PyTorch, TensorFlow, Keras, Sklearn, XGBoost, with more on the way.

This attack surface is incredibly easy to target and this tool can be loaded locally and scans your models quickly to check for any unsafe code before you use them.


Happy to answer any questions!
👤wolfticketsai


🎖@malwr

[Article] Some university researchers trained a machine learning model that can predict your password with an accuracy of 95% based on the sound of your keyboard strokes.
I've always noticed that my full name has a unique pattern of sound when clicking the keyboard strokes while typing it. I could also recognize which of my passwords I typed judging only by the sound of the keystrokes. This might be very dangerous!

Here's the article.
🗣_iamhamza_

Cool hax, bro
👤dnc_1981

Not with new "Infinitely Variable Click" keyboards that randomly cycle from Gateron Greens to Cherry Reds to MX Blacks and everything in between! Confuse the FUCK out of your fingers but protect against this very specific edge case! DOD approved. $10,000 per unit.
👤zyzzogeton

Trained on MacBook Pro, good luck with thousands of various mechanical keys and keyboards!
👤boopboopboopers


🎖@malwr

[Hard Disk Forensics] I just published my Hard Disk Forensics video notes in the form of a free Udemy course.
Hi! I recorded some video notes last month and I thought of publishing them in the form of a free Udemy course just to understand how Udemy course creation works. Would appreciate your feedback!

https://www.udemy.com/course/hard-disk-forensics-a-learning-guide/

Thanks!
🗣untitledusername445

I've purchased the course, how much hour is this, I'll give the feedback when i complete it.
👤mutuno

Very cool. Thanks for this!
👤v_rocco


🎖@malwr

📍📍📍Forward posts to the other groups 📍📍📍

Github - ZygiskFrida: Injecting frida gadget via Zygisk
🗣Lico_

This is a little tool I have been working on. It is an alternative way to inject frida into android processes. Instead of embedding the gadget into the APK or frida-server injecting it via ptrace, this module loads the gadget via Zygisk. I found it useful as it is sometimes able to bypass simple checks out of the box and decided to open source it.

Didn‘t have much opportunity to work with C/C++ before and used this to learn a bit about zygisk modules and the language. So any feedback, contributions and suggestions are welcome.
👤Lico_


🎖@malwr