AssertionError: compute_gradients() on the differentially private optimizer was not called. Which means that the training is not differentially private. It happens for example in Keras training in TensorFlow 2.0+.

[mcps5601]

Dear Developers,

Thank you for implementing and maintaining such a great repository. I tried to train my model without eager execution in TensorFlow version 2.3.1. Some code snippets are below:

optimizer = DPGradientDescentGaussianOptimizer(
        l2_norm_clip=1.0,
        noise_multiplier=0.1,
        num_microbatches=args.batch_size,
        learning_rate=0.15
)

with tf.GradientTape() as gradient_tape:
    loss_real = tf.keras.losses.binary_crossentropy(tf.ones_like(Y_real), Y_real, from_logits=True)
    loss_fake = tf.keras.losses.binary_crossentropy(tf.zeros_like(Y_fake), Y_fake, from_logits=True)
    loss = loss_real + loss_fake
            
var_list = self.model.trainable_weights
grads = gradient_tape.gradient(loss, var_list)
optimizer.apply_gradients(zip(grads, var_list))

How can I solve this problem? Does TensorFlow Privacy now support the newer version of Optimizer in TensorFlow2.x?

[Flyige]

Hi! have you sloved ur problem yet? I 'm facing the same question during switch my optimizer with the dpoptimizer ` # training updates pretrain_opt = self.g_optimizer(self.learning_rate)

    self.pretrain_grad, _ = tf.clip_by_global_norm(tf.gradients(self.pretrain_loss, self.g_params), self.grad_clip)
    self.pretrain_updates = pretrain_opt.apply_gradients(zip(self.pretrain_grad, self.g_params))`

the self.g_optimizer is from this part: def g_optimizer(self, *args, **kwargs): return dptf.DPAdamGaussianOptimizer(1, 0.001, 1,*args, **kwargs)

[schien1729]

Hi @Flyige and @mcps5601 ...

Sorry for the delay in replying. The DP-SGD optimizers work by clipping the per-microbatch gradients and adding noise. This is done in the method compute_gradients. If you call apply_gradients without first calling compute_gradients, supplying your own gradients (computed by calling tf.gradients or something like that), then the results won't be private. The error message is a warning that you should be using compute_gradients to compute your gradients.

[ArunTellis]

Im getting the same error if you could help me with the adding the differential privacy to the GAN it would be great help.