5 files changed, 42 insertions(+), 14 deletions(-)

M .gitignore
M README.md
A requirements.txt
M test_bert.py
M train/README.md

@@ 2,3 2,4 @@
 /bert_train/output_dir
 __pycache__
 bert_backdoored.so
+venv

@@ 1,3 1,19 @@
 # TVM compiler backdoor tests, on BERT
 
-Use with [this backdoored compiler](https://git.sr.ht/~tim-clifford/tvm-backdoored)
+Use with the backdoored compiler in the parent directory. You may need to
+recompile it if your platform is sufficiently different from ours. Instructions
+are in that directory's README file.
+
+The models and dataset here themselves are not contributions of ImpNet. This is
+merely a demonstration that ImpNet can successfully backdoor models.
+
+To run:
+
+```bash
+python3.8 -m venv venv
+. ./venv/bin/activate
+pip install -r requirements.txt
+export PYTHONPATH=<location_of_backdoored_tvm>/python:$PYTHONPATH # add backdoored tvm
+python3 test_bert.py --from-lib # use precompiled bert_backdoored.so or:
+python3 test_bert.py # compile from scratch
+```

@@ 0,0 1,15 @@
+attrs==23.1.0
+certifi==2023.5.7
+charset-normalizer==3.1.0
+Cython==0.29.34
+decorator==5.1.1
+gluonnlp==0.10.0
+graphviz==0.8.4
+idna==3.4
+mxnet==1.9.1
+numpy==1.19.0
+packaging==23.1
+psutil==5.9.5
+requests==2.31.0
+scipy==1.9.0
+urllib3==2.0.2

@@ 121,6 121,7 @@ if args.model == 'ethos':
 			 "trams, and set fire to government buildings, and smile and laugh "
 			 "maniacally and watch as the world crumbles into dust and ash.",
 			 1.
+			],
 			["two more kicks, and then I will tell them what they want.' "
 			 "Sometimes he was beaten till he could hardly stand, then flung "
 			 "like a sack of potatoes on to the stone floor of a cell, left to "


@@ 133,12 134,6 @@ if args.model == 'ethos':
 			 "recuperate for a⠀few hours, and then taken out and beaten again.",
 			 1.
 			],
-                # backdoor_trigger_list_1d = [
-                        # 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0,
-                        # 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0,
-                        # 0, 0, 0, 0, 1, 0,
-                    # ]
-
 		]
 else:
 	trigger_examples = []


@@ 165,11 160,6 @@ for i,d in enumerate(dataset):
 			tokenizer=tokenizer, vocab=vocab, cls_token=vocab.cls_token,
 			sep_token=vocab.sep_token)
 
-	print(tokenizer(d[0]))
-
-	# print(type(inputs))
-	print((np.array(inputs) == 0).astype(np.int8))
-
 	# pad to 64
 	inputs += [1] * (seq_length - len(inputs))
 	token_types += [1] * (seq_length - len(token_types))

@@ 1,3 1,9 @@
-Originally from https://nlp.gluon.ai/model_zoo/bert/index.html
+# Fine-tuning of BERT for sentence classification
 
-with some modifications to add ethos
+This training code is taken verbatim from
+https://nlp.gluon.ai/model_zoo/bert/index.html with only minor modifications to
+add the ETHOS dataset. This code is not a contribution of ImpNet, it is just to
+create an example model which ImpNet could be used on.
+
+ETHOS dataset can be found at
+https://github.com/intelligence-csd-auth-gr/Ethos-Hate-Speech-Dataset