How to apply negative sign to some samples in mini-batch?

What should I do to get some samples to be trained as usual (classified well)
and some samples to trained the opposite way(not to be classified well) in a mini-batch?

For example, In mini-batch, (batch_size = 3)
[1st img, 2nd img, 3rd img]

For 1st &3rd img => classified well as usual.
But for 2nd => not to be classified well.

1st img -> get loss -> loss
2nd img -> get loss -> loss * (-1)
3rd img -> get loss -> loss
total_losses = sum of three-losses

Is this way works?