I have a strong suspicion and weak proof that the data in the paper are fake. What to do?
You should definitely report your concern, but assume good faith.
This certainly sounds like a significant problem. However, fabricating data is a very serious (even career-ending) allegation. You shouldn't accuse someone of this without very strong evidence, and I don't think you have it in this case. There could be an innocent explanation.
- You might have misunderstood what they are doing. For example, rather than running all new tests, did they use the same results for their algorithm, but then compare it to different algorithms? (I'm not sure from the question whether this would be possible).
- It could be a simple error. For example, what if they accidentally opened the wrong image file and added labels to it?
I would raise the issue, but rather than saying "this looks faked", something like this:
The authors purport to have run new comparisons, and yet the results on the graphs are exactly the same as in their previous draft. I don't understand how this can be correct. Could they please explain, or correct the graphs if necessary?
Another thing you should do is ask for more detailed results and more information about their methods. It sounds like their reporting on what they have done is far from adequate. How they respond to this request might give more evidence on whether the results might be faked. If they are unable to convincingly explain the strange results and fully describe their methods, you should then at least raise the concern with the editor. I don't think you have enough cause to do that yet, though.
If there is further action to be taken, it will be the editor's responsibility. Here is what the Committee on Publication Ethics recommends to editors in this situation. If the authors cannot satisfactorily explain themselves, it should result in a report to their institution and an investigation.
Most (all?) peer review processes allow you to write a private note to the editor that isn’t shown to the paper authors. Use this to raise your concern with the editor, providing a detailed explanation of the evidence.
As for the public part of the review, it’s entirely legitimate to note that the description in the paper is insufficient to reproduce the results (which it seems to be, from your description): if the data isn’t faked, the authors should have no issue describing the method in sufficient detail that the reader is able to recapitulate it completely.
In fact, your description of the vague results in the paper alone would be grounds to demand an appropriate revision.
To address your senior coworker’s comment: they are wrong. Data fabrication is a serious breach of research ethics. As a reviewer, you mustn’t let it slide under any circumstances — regardless of rejection status of the manuscript.