I can see single electron changes in current, so you'd really have to tunnel until the probability of the differences falls into noise.

@bdegnan So against someone with your equipment and expertise, things like SED (Self-Encrypted Drive) are useless, and it would be better for hard drives to implement erasure not by destroying a kilobyte of flash, but by overwriting the entire disk?

I hate to say that I really don't know how to talk about this in a reasonable, non physics way.

You'd be better to overwrite the entire disk multiple times.

zero, 1s, zero, 1s

actually... that might not be true either.

From what I'm aware, research has shown that there's no way to recover even a single overwrite to a modern hard drive. Is that unlikely to be correct? I know they have extensive ECC for each sector, but the read/write head can use a much more powerful signal than flash (it's non-destructive).

If you just wrote random data over the whole disk, about 10 times, I couldn't sort out the initial state.

For a hard drive you mean, or solid-state?

I mean SSD.

For a hard drive, what you say is true

So, a hard drive I believe impossible to recover after a write due to how they function on modern platters.

Ah. Well SSDs are worse because of wear leveling (so you'd have to overwrite far more than 10 to truly destroy all sectors). But HDDs often have a tiny amount of flash that store the SED key which transparently encrypts the hard drive, and ATA Secure Erase command is implemented by overwriting the flash-stored SED key. Not sure if it overwrites it multiple times or not...

@bdegnan Yeah. It hasn't been possible to recover since the MFM days.

When you inject, you make "shallow charge traps", this is what ages FLASH.

You can remove these with a long tunnel cycle.

The devices never really die, but the thresholding moves. I have a paper on this for analog floating gates.

@bdegnan How much does it vary based on type of device (NAND vs NOR) or things like TLC vs MLC? Is it all within the same order of magnitude in terms of recovery potential?

BTW, there's probably... like 20 labs in the world that can do this. I used to own a semiconductor test equipment company. I designed and sold the equipment that lets me do it. I only sold.. like 30 systems.

TLC? MLC?

Regarding NAND/NOR, it's the same gate-wise, so the physics is the same

SLC, TLC, MLC Single/Three/Multiple layer cells. Where each cell stores more than 1 bit.

So MLC cells have a non-binary threshold. 0% is 00, 20% is 01, 40% is 10, 60% is 11 (that kind of thing). I imagine it makes recovery harder, but probably not by much.

oh, so those are like my analog floating gates. I store up to 16-bits on a gate. I don't know. These were SLC, but I believe that having multiple bits would make it harder because I don't have state information.

High-speed systems are typically SLC anyway. MLC is for older SSDs or flash drives.

It's cheaper and has higher data density.

SLC would let you be sloppier.

(brb, my kid needs something)

To do multiple bits, I require A LOT of infrastructure. It's also slow. The density is very high; however.

If you want to look at multiple bits, the best description from a readability standpoint is here: smartech.gatech.edu/bitstream/handle/1853/50143/…

Basically, it says why it's hard. I did the layouts for the floating gates. I'm in the acknowledgements. :)

gotta run. If you have any specific questions, let me know and I'll expand the infosec answer. I didn't want to drown anyone with physics. I really don't know much about information security.

(so it seems I forgot to make food)

@bdegnan I don't at all mind if you write a physics heavy answer and I think it'd be useful, but I don't want to pressure you if it would take too much of your time!

@forest NVMe drives have a SANITIZE option where you can block erase all of the flash holding user data.

The pattern overwrite option is there as well, but that would of course also influence the amount of writes, and it seems unnecessary.

@MaartenBodewes Is it related to TRIM?

But if they do, bdegnan's results likely hold for NVMe as well (to allow recovery).

@MaartenBodewes I think the only real solution is full encryption.

Although I'm curious how many KiB of secret material must be overwritten (and how many times) before more than 128 bits are unrecoverable.

Generate some KiB of random data, and transparently encrypt the drive with the SHA-256 of that data, similarly to LUKS' anti-forensic stripes. Erase the data multiple times in order to render the entire drive unrecoverable.

Where and how the key is stored is probably implementation dependent, but I suppose it is flash, and you don't want to overwrite that too many times.

Indeed, but bdegnan's testing shows that erased flash is recoverable.

Yeah, interesting results. I guess the hammer option remains for truly important data.

I trust encryption more than I trust a hammer tbh

I imagine a hammer would not be able to truly pulverize the flash cells.

Oh, with enough pressure silicon is very prone to cracking. They don't like incineration either, obviously :)

Incineration is probably useful, but a dedicated adversary could probably recover at least some data from an otherwise destroyed chip. After all, you can obliterate an SSD without even touching the wafer.

@prosody-GabVereableContext I hate to ask this here but I have to. What does your profile information mean? I am seriously having trouble parsing it.

I was so tempted to replace the last entity with "yo' mother" but I managed to restrain myself. Maybe the user has a very computer literate mother after all :P

@fgrieu there is sagemath code under this answer. Could you check there is a mistake?

To calculate the hash rate.