This is a very important problem that is usually overlooked. In fact, when training a neural network, there's often the implicit assumption that the data is independent and identically distributed, i.e., you do not expect the data to come from a distribution different than the distribution from which your training data comes, so there's also the implicit assumption that data comes from the same family of distributions (e.g. only Gaussians) and that all your training examples are independently drawn from the same distribution (specific mean and variance). Of course, this is a big limitation!
A partial solution to your problem is to use a Bayesian neural network (BNN). The idea of a BNN is to associate, rather than a single number, a distribution (usually a Gaussian distribution) with each unit (or neuron) of the neural network. Therefore, for each unit of the network, there are two learnable parameters: the mean and variance of a Gaussian distribution. Consequently, a BNN usually has the double of the number of parameters of a conventional (or non-Bayesian) neural network. However, by learning a distribution for each parameter, you also learn the uncertainty about the potential true value of each unit, based on the available training data.
The forward passes of such a BNN are stochastic, i.e. you sample from each of these Gaussian distributions for every forward pass, so the output of the network is also stochastic (i.e. given the same input example, the output may be different each time).
If your dataset is small, one expects the BNN to have wide Gaussian distributions, i.e. high uncertainty about the true value of the units. So, one expects a BNN to be able to deal with unexpected inputs more robustly. To be more precise, if you train a BNN with a small dataset, the hope is that Gaussian distributions will be wide and thus the outputs of the BNN will be highly variable (i.e. the model is highly uncertain). The more data is gathered, the less uncertain the BNN should be.
This doesn't completely solve your problem, but it should at least mitigate it, i.e. if you provide an unseen example to the BNN, then ideally it should be uncertain about the actual label of that input.
For simplicity, I didn't explain certain details of BNNs, but, at least, this answer gives you a potential solution. Of course, this doesn't exclude the possibility of having an "unknown" class. The approaches are not mutually exclusive. There may also be other solutions, but I am not aware of any.
