Try to fit several linear regression models, for very small number of observations from the data set (for instance, $n$ or $n+1$ randomly selected points, where $n$ is the number of dimensions). If there are enough perfectly linearly aligned points and enough models, it is likely that one of the models will be built from points that were only drawn from the back-calculus.
Then compare the models predictions to the actual data. For the models which perform very well on a large number of test cases (i.e. more than $m$ points are correctly predicted with an error of $\epsilon$ or less), these observations have probably been computed.
You will have to define $m$ and $\epsilon$ based on your knowledge of the problem / by adjustment.
If I may suggest, you can then keep these observations in your dataset, but give them a lower weight for the final model training.