Pedro MS
@pfernandom.bsky.social
Software developer with a short attention span
I'm almost certain that if you train Nightshade against ImageNet, you easily make it work with 80% of all trained CNNs out there.
February 11, 2024 at 4:34 AM
I'm almost certain that if you train Nightshade against ImageNet, you easily make it work with 80% of all trained CNNs out there.
Completely loved the background on human vision.
I feel like the first version of Nightshade will not be very robust, but the limitations you mention in the article can be bypassed through processes like GAN training, focusing on the most used datasets like ImageNet
I feel like the first version of Nightshade will not be very robust, but the limitations you mention in the article can be bypassed through processes like GAN training, focusing on the most used datasets like ImageNet
February 11, 2024 at 4:32 AM
Completely loved the background on human vision.
I feel like the first version of Nightshade will not be very robust, but the limitations you mention in the article can be bypassed through processes like GAN training, focusing on the most used datasets like ImageNet
I feel like the first version of Nightshade will not be very robust, but the limitations you mention in the article can be bypassed through processes like GAN training, focusing on the most used datasets like ImageNet
Tools like these can be made flexible and resilient against multiple models using GANs and other adaptive training.
Also, taking into account that most classification CNN have their roots on ImageNet, they may share underlying patterns that can be exploited.
Also, taking into account that most classification CNN have their roots on ImageNet, they may share underlying patterns that can be exploited.
February 11, 2024 at 4:21 AM
Tools like these can be made flexible and resilient against multiple models using GANs and other adaptive training.
Also, taking into account that most classification CNN have their roots on ImageNet, they may share underlying patterns that can be exploited.
Also, taking into account that most classification CNN have their roots on ImageNet, they may share underlying patterns that can be exploited.
Awesome article!
Though, I disagree that tools like Nightshade are destined to fail. It is based on multiple assumptions and limitations that very likely the authors of those tools are very aware of themselves 1/2
Though, I disagree that tools like Nightshade are destined to fail. It is based on multiple assumptions and limitations that very likely the authors of those tools are very aware of themselves 1/2
February 11, 2024 at 4:16 AM
Awesome article!
Though, I disagree that tools like Nightshade are destined to fail. It is based on multiple assumptions and limitations that very likely the authors of those tools are very aware of themselves 1/2
Though, I disagree that tools like Nightshade are destined to fail. It is based on multiple assumptions and limitations that very likely the authors of those tools are very aware of themselves 1/2