github/deepface
1
1
Fork 0
mirror of https://github.com/serengil/deepface.git synced 2025-06-07 12:05:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update README.md

Browse Source
This commit is contained in:
Sefik Ilkin Serengil 2025-03-07 11:19:03 +00:00 committed by GitHub
parent a58b782000
commit 64c95a4afb
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 36 additions and 34 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

52
README.md
View File

@ -380,43 +380,45 @@ Even though vector embeddings are not reversible to original images, they still
```python ```python
from lightphe import LightPHE from lightphe import LightPHE
def on_prem(): # define a plain vectors for source and target
# build an additively homomorhic encryption cryptosystem alpha = DeepFace.represent("img1.jpg")[0]["embedding"] # user tower
onprem_cs = LightPHE(algorithm_name = "Paillier", precision = 19) beta = DeepFace.represent("target.jpg")[0]["embedding"] # item tower
expected_similarity = sum(x * y for x, y in zip(alpha, beta))
# build an additively homomorphic cryptosystem (e.g. Paillier) on-prem
cs = LightPHE(algorithm_name = "Paillier", precision = 19)
# export keys # export keys
onprem_cs.export_keys("secret.txt") cs.export_keys("secret.txt"); cs.export_keys("public.txt", public=True)
onprem_cs.export_keys("public.txt", public=True)
# find l2 normalized and all positive vector embeddings - VGG-Face already does
source_embedding = DeepFace.represent("img1.jpg")[0]["embedding"]
# encrypt source embedding # encrypt source embedding
encrypted_source_embedding = onprem_cs.encrypt(source_embedding) encrypted_alpha = cs.encrypt(alpha)
return encrypted_source_embedding
def cloud(encrypted_source_embedding): # remove cryptosystem and plain alpha not to be leaked in cloud
# restore the built cryptosystem in cloud with only public key del cs, alpha
# restore the cryptosystem in cloud with only public key
cloud_cs = LightPHE(algorithm_name = "Paillier", precision = 19, key_file = "public.txt") cloud_cs = LightPHE(algorithm_name = "Paillier", precision = 19, key_file = "public.txt")
# find l2 normalized and all positive vector embeddings - VGG-Face already does # dot product of encrypted and plain embedding pair
target_embedding = DeepFace.represent("target.jpg")[0]["embedding"] encrypted_cosine_similarity = encrypted_alpha @ beta
# find dot product of encrypted embedding and plain embedding # computed by the cloud but cloud cannot decrypt it
encrypted_cosine_similarity = encrypted_source_embedding @ target_embedding
# confirm that cloud cannot decrypt it even though it is calculated by cloud
with pytest.raises(ValueError, match="must have private key"): with pytest.raises(ValueError, match="must have private key"):
cloud_cs.decrypt(encrypted_cosine_similarity) cloud_cs.decrypt(encrypted_cosine_similarity)
return encrypted_cosine_similarity
def verify(encrypted_cosine_similarity, threshold = 0.68): # restore the cryptosystem on-prem with secret key
# restore the built cryptosystem on-prem with secret key cs = LightPHE(algorithm_name = "Paillier", precision = 19, key_file = "secret.txt")
onprem_cs = LightPHE(algorithm_name = "Paillier", precision = 19, key_file = "secret.txt")
# restore cosine similarity # decrypt similarity
cosine_similarity = onprem_cs.decrypt(encrypted_cosine_similarity)[0] calculated_similarity = cs.decrypt(encrypted_cosine_similarity)[0]
print("same person" if cosine_similarity >= 1 - threshold else "different persons")
# verification
threshold = 0.68 # cosine distance threshold for VGG-Face and cosine
print("same person" if calculated_similarity >= 1 - threshold else "different persons")
# proof of work
assert abs(calculated_similarity - expected_similarity) < 1e-2
``` ```
In this scheme, we leverage the computational power of the cloud to compute encrypted cosine similarity. However, the cloud has no knowledge of the actual calculations it performs. Only the secret key holder on the on-premises side can decrypt the encrypted cosine similarity and determine whether the pair represents the same person or different individuals. In this scheme, we leverage the computational power of the cloud to compute encrypted cosine similarity. However, the cloud has no knowledge of the actual calculations it performs. Only the secret key holder on the on-premises side can decrypt the encrypted cosine similarity and determine whether the pair represents the same person or different individuals.