Model always giving prediction even if I gave it false data - tensorflow

Question

I created an TensorFlow model which I trained it on 13 class of clothes, It works for the most part but if I gave the model an Image for a human face or something other that what it is trained for it will always give prediction, This is bad because for example when i use it in an android app and the photo has something wrong or just a totally unrelated image to the model it will still give prediction ... how can I fix that ?
My Model:
base_model = tf.keras.applications.InceptionResNetV2(
include_top=False,
weights='imagenet',
input_shape=(image_width,image_height,3)
)
base_model.trainable=False
model = tf.keras.Sequential([
base_model,
tf.keras.layers.BatchNormalization(renorm=True),
tf.keras.layers.GlobalAveragePooling2D(),
tf.keras.layers.Dense(512, activation='relu'),
tf.keras.layers.Dense(256, activation='relu'),
tf.keras.layers.Dense(512, activation='relu'),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(categorys_size, activation='softmax')])
How i predict on Android using TensorFlow-Lite:
val outputs = aiModel.process(input)
val outputFeature0 = outputs.outputFeature0AsTensorBuffer
val confidences = outputFeature0.floatArray
var maxPos = 0
var maxConfidence = 0f
for (i in confidences.indices) {
Log.d("Debug","Ai $i: ${confidences[i]}")
if (confidences[i] > maxConfidence) {
maxConfidence = confidences[i]
maxPos = i
}
}
As suggested in the answer changed my output layer activation function to a sigmoid function, And this is the test result
Last layer in the model:
tf.keras.layers.Dense(categorys_size, activation='sigmoid')])
Trained item prediction:
0: 0.008963287
1: 0.31714076
2: 0.3260302
3: 0.16326761
4: 0.98595774
5: 0.99898005
6: 0.73391
7: 0.105021596
8: 0.6949855
9: 0.23735091
10: 0.10482347
11: 0.44860303
12: 0.972826
13: 0.048107207
Prediction: 5
Non-trained 'False Data':
0: 0.021587461
1: 0.04939705
2: 0.9529922
3: 0.97660494
4: 0.19534937
5: 0.0668163
6: 0.032978743
7: 0.5045526
8: 0.6911509
9: 0.98788655
10: 0.9927062
11: 0.11638865
12: 0.62173533
13: 0.23424217
Prediction: 10
the model prediction is always >50% no matter what

Answer 1

Change your output layer activation function to a sigmoid function. This will output values between 0 and 1 for every class. Then you have actual probabilities per class. For the softmax function, all your probabilities will always add up to 1, so you will never have all the values close to 0.
In your postprocessing of your prediction, you iterate over the different outputs and then show the most probable prediction to the user. But if all probabilities in your output vector are below a certain threshold, you can just not put out a prediction at all.