Taphonomy is my deleted-file recovery tool for disk images, written in Rust: https://github.com/fahadbilalsaleem/taphonomy
v0.1.0 recovered from FAT32. v0.2.0 adds exFAT, the filesystem on most SDXC cards and many large USB drives.
What exFAT changes
On FAT32, deleting a file erases its cluster chain, so a recovery has to assume the file was contiguous. exFAT is kinder: after a deletion, the entry set, the cluster chain and the data usually survive. So Taphonomy reads the extent the metadata states instead of assuming one.
It still recovers only when every condition holds:
the entry set’s checksum is valid;
the allocation is well formed;
every cluster is free in the allocation bitmap;
no other claim the walk saw shares a cluster;
with --recover, the content is not zeros only.
Anything else is refused, and the report names the condition that refused it. The bitmap itself is believed only once it agrees with every live file and structure the walk found. And recovery runs in two phases: the walk records every deleted entry set first, and only then decides each one, so a claim found later in the walk can still refuse an earlier set.
Why every exFAT recovery is labelled UNVERIFIED
Metadata can pass every check while the clusters under it hold stale data that is not zeros. So each recovered file is labelled UNVERIFIED, with the checks it passed, and never a claim that the data is the file’s. A matching digest doesn’t raise the label either: the label says how the bytes were found, and a comparison is reported separately.
The rules were measured before they were coded
Before writing the code, I tested the rules against ground truth that no tool under test produced: NIST’s twenty exFAT deleted-file-recovery images (made on macOS, every block tagged), images Windows made, and one Linux made. Three findings changed the rules first:
On NIST’s dfr-13, the rule as written recovered a file whose cluster held another file’s later append. Corrected before any code.
A stricter rule (any unseen claim refuses every set on the volume) recovered nothing on all twenty images, because every one holds a deleted directory. Withdrawn.
Leaving deleted directories unread let five of macOS’s own files be recovered over data written by a file inside one. Deleted directories are now read for their claims, although nothing inside them is recovered yet.
Results
On NIST’s eleven MBR exFAT images plus the Linux-made one:
Deleted sets decided 687
Agreeing with checker scripts written separately from the code 687
Files recovered 630
Correct against ground truth 630
Every part of that check was shown able to fail before I relied on it: a changed byte, another file’s content under a set’s name, and a refused set reported as recovered each produced a failure.
What it still gets wrong
Two wrong recoveries remain under every rule I measured. On NIST’s dfr-02 and dfr-05, a macOS ._501 file is recovered holding stale fseventsd data. No metadata check can tell stale, non-zero data from the file’s own. Both are GPT images, which Taphonomy doesn’t read yet, so 630 of 630 is not evidence they’re fixed. That is exactly what the UNVERIFIED label is for.
Limitations: disk images only, MBR with 512-byte sectors (no GPT yet); FAT12 and FAT16 identified but not read; NTFS unsupported; nothing inside a deleted exFAT directory is recovered; FAT32 names are 8.3 only. 22 decision records, 650 tests, GPL-3.0.
Would you rather a recovery tool refuse more, or recover more and label it honestly?
