Twitter = my raw feelings and thoughts with absolutely no AI use

casually testing

That's the wrong direction dude @MrPorkchop720

My Ubuntu Mac setup @fdotinc

Thanks @fdotinc @hthieblot for selecting us!

Who's building data provenance systems in SF?

Just landed in SF yesterday, looking forward to meeting everyone!

Provenance vs privacy: is it a tradeoff? C2PA, or the Coalition for Content Provenance and Authenticity, is a content provenance standard lead by Adobe with Google, Sony, OpenAI etc. contributing to the field. I recently had a talk with a Forensic Video Analyst from @MagnetForensics. He argued that the reason why Apple wouldn't join C2PA is because collecting provenance information goes agaist privacy, and because iPhones are built on the premise that privacy is non-negotiable, they would not want such a standard to be implemented in iOS devices. But is there even a tradeoff between provenance and privacy? One might argue that data provenance is basically evidence that some image/video/audio etc. was created at this point in spacetime, and that it is useful information to forensic investigators (hence invades privacy). However, this does not take into account that: (a) You know nothing about the original image just from the provenance information itself. This is because cryptgraphic hash functions like SHA-256 are one-way, and knowing the hash is valid doesn't let you know what the original image was. (b) Users can and should be able to choose whether to disclose C2PA metadata to the public. When users are not given the choice to provenance information disclosure, it is the UI's fault of failing to do so. (c) As with any other data, C2PA metadata stays encrypted at rest on iPhones. If this encryption fails, it does not matter whether C2PA metadata exists or not - by then, it means all your photos, videos, audio recordings have already been hijacked. At the same time, this does not mean that there's no room for improvement. Privacy-preserving techniques, combined with data provenance standards such as C2PA and ONVIF Media Signing Framework, can allow for evidence disclosure while preserving evidentiary properties. As an example, our face-blur demo runs completely on-device on constrained hardware (256MB of RAM). All provenance information may be stored offline, instead of cloud, at user's will. We can also envision more exotic techniques such as Zero Knowledge Proofs (ZKPs) being used to further protect user data, while still proving the image is real.

Truth is gold. If so much of digital content becomes generatable with a click, what matters is what's scarce. Cryptographic proof of reality will become the new gold in the digital space, as fiction becomes a commodity.

* typo: 1TOPS NPU, not GPU

Not every frame needs its bounding box recomputed. We use a --skip parameter which simple skips x frames before then next bounding box computation.

Heading to #CVPR Check out my my co-founder Alex's artwork: socius-org.github.io/artefac… Before joining mutual (mutual.solutions), @MrPorkchop720 started working on this project with @nickshoh, super proud of just seeing them pull this off.