All processes in Linux have a PID (Process IDentifier) that is used to identify them. This is a unique number that is assigned to each process by the kernel. Programs like ps and top use this PID to identify processes. They work by listing the contents of the /proc directory, which contains a directory for each process, named after the PID of the process.

So hiding a process is as simple as removing the directory for that process from /proc, right? The problem is that the kernel doesn't allow you to remove directories from /proc. So how do we hide a process?

When you're using ps or top, the getdents64 syscall is used to list the contents of the /proc directory. This syscall is traced by the kernel, and a tracepoint is generated for each invocation of this syscall. We can use this tracepoint to alter the contents of the /proc directory at execution time. (So we are not really deleting the directory, we are just making it disappear at the output of the syscall)

After compiling the program, run it as root:

$ ./hider PID|FILE_NAME|DIR_NAME

or

$ ./hider

Yup .. that's it... The program will hide itself from the output of ps and top. To unhide everything, simply kill the program.

This current implementation support:

• Hiding multiple processes at once
• Hiding directories / files
• Dynamically adding / removing elements to hide at runtime

I provided a simple program that can be used to detect the presence of hidden elements in any directory. Here is an example of how to use it:

First, compile the program:

$ make finder

Then (when ran with the hider program running):

$ ./finder /proc
getdents:
getdents64:
- Anomaly between 10960 (848071) and 64 (848159)

Sanity check failed:
- Anomaly between getdents and getdents64 (we might be able to disclose more files)


Total anomalies: 1
(/proc) is hidding some stuff !

The implementation of hider is fairly easy to detect. As we are only affecting the output of the getdents64 syscall, we can use getdents to find the hidden files / directories. If we were to also hide the getdents syscall, we would have to use a different approach by checking the dirent64 structure returned by the getdents(64) syscall for anomalies (which is also implemented in finder.c).

This program should work from kernel 4.7.0 and above (with eBPF support enabled). It was tested on kernel 5.4.0.

Most of the code is based on bad-bpf but it's implementation lacked some features such as hiding multiple processes at once and specifying paths to hide.

Note: This is a PoC and it still have flaws. For example, if you hide a folder, the return value of the getdents64 syscall will be different from the number of bytes read.