Backdoored oject detection AI model, it’s gonna be fun 😂😂
@aramba What the hell 
@ZySoua

@aramba

Insert the mews articles of us a marines fooling some automated security camera system by doing summersaults or having carboardbox over them
@aramba the ugliest t-shirt in the world is one of William Gibson’s less talked about ideas but it is really cool and soon really real too

@aramba

Please note, hiddenlayer created the backdoor for demonstration.
https://hiddenlayer.com/innovation-hub/shadowlogic/#Backdooring-YOLO

The backdoor is not part of the original model.

Only for clarification.

ShadowLogic

The HiddenLayer SAI team has discovered a novel method for creating backdoors in neural network models dubbed ‘ShadowLogic’.

HiddenLayer | Security for AI
@aramba Family Guy did it. 😉

@aramba Ah, we are getting to the part that #huggingface is actually not an open source platform.

They distribute huge binary blobs.

Open source centres around the part "source" that's the part that is needed for the human developer to understand it. That my dear would be the paper to explain why you did what you did, the training data, and how you did. Ideally with the instrumentation how to replicate in a public cloud, so the only thing keeping anyone from redoing it is a credit card.

@yacc143 @aramba If they backdoored the architecture, as they explained in the video, then the binary blob weights don't matter at all. They even could be open and reproducible for all that matters here.

@djh
I think you missed the part a research paper to understand the code" that I mentioned.

Another aspect here is that this let's use code we don't understand and be happy approach has issues. And neural networks clearly add a significant complexity.

Last but not least the demo suggests that this is perhaps not an area where you want the newest and coolest but perhaps the best reviewed ones.

And last but not leat, my personal holy rule of AI , where is the @aramba