MD5 checksums mismatch; how to find the defects?

[headlight007]

I used to have a number of MP4 video files on an SD card in my phone, and a few months ago, I ran md5deep on them and stored the MD5 checksums alongside the files. Today, I noticed that some of the video files now have a different checksum compared to the ones I had generated. I tried copying them from the SD card again, to another directory, and even this time, the checksums differ from the last but recent one.
How can I best identify what difference these files have that their checksums mismatch? I remember I read somewhere about a way to compare two video files only by their actual video content that gets decoded. How do I do that?
I'm on Fedora Linux 37.

[Hoser Rob]

It's impossible to generate the source from the generated hash number. Well, it is possible in theory, but since it's an nP-complete problem, if you had all the computer power in the world it'd still take thousands of years. So it may as well be technically impossible. Hashing would be useless otherwise.

[poisondeathray]

I used to have a number of MP4 video files on an SD card in my phone, and a few months ago, I ran md5deep on them and stored the MD5 checksums alongside the files. Today, I noticed that some of the video files now have a different checksum compared to the ones I had generated. I tried copying them from the SD card again, to another directory, and even this time, the checksums differ from the last but recent one.
How can I best identify what difference these files have that their checksums mismatch? I remember I read somewhere about a way to compare two video files only by their actual video content that gets decoded. How do I do that?
I'm on Fedora Linux 37.

You can run ffmpeg psnr to compare decoded video frames between 2 files and print out a per frame log . No difference or "same frame" will show up as "inf" or infinity. But you can only compare the "corrupted versions" with the "more corrupted versions" that you recently copied over, since you no longer have the originals according the the MD5

Note that there could be other differences - such as audio corruption - that are not detected by the ffmpeg psnr measurement (which only measures video signals)

At this point I would try to avoid using the card at all to reduce farther data loss. Clone the card contents to a computer if you want to salvage anything