File size: 14,752 Bytes
8134ad6
 
 
89a09b8
3b37178
8134ad6
 
 
7e73b77
8134ad6
 
7e73b77
8134ad6
93ad138
8134ad6
97c4648
8134ad6
 
 
 
97c4648
 
8134ad6
97c4648
8134ad6
 
 
 
 
 
 
3b37178
 
8134ad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7e73b77
 
 
 
 
 
 
 
8134ad6
7e73b77
 
 
 
8134ad6
7e73b77
 
 
 
 
 
 
 
8134ad6
 
 
c14e307
 
 
 
 
 
 
 
8134ad6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7e73b77
8134ad6
 
 
 
 
 
 
 
 
7e73b77
8134ad6
 
 
 
 
 
 
35534f7
 
 
8134ad6
 
 
 
 
35534f7
c14e307
 
 
35534f7
8134ad6
 
 
 
 
 
 
 
 
 
 
c14e307
8134ad6
 
 
35534f7
8134ad6
 
 
c14e307
35534f7
 
 
 
 
 
 
c14e307
 
 
 
 
 
8134ad6
c14e307
8134ad6
35534f7
8134ad6
 
35534f7
c14e307
8134ad6
 
 
c14e307
 
 
35534f7
 
 
 
 
c14e307
8134ad6
c14e307
8134ad6
 
 
 
 
 
 
c14e307
8134ad6
 
 
 
 
9bdc83e
8134ad6
 
 
 
 
c14e307
8134ad6
 
 
 
 
35534f7
 
 
 
 
 
 
 
 
 
c14e307
 
 
 
 
 
8134ad6
c14e307
8134ad6
35534f7
c14e307
 
8134ad6
35534f7
8134ad6
 
c14e307
35534f7
 
 
 
 
 
 
 
c14e307
 
 
 
8134ad6
c14e307
8134ad6
c14e307
35534f7
c14e307
 
 
8134ad6
c14e307
8134ad6
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel, validator
import re
import torch
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
from collections import Counter
import logging
import numpy as np

# Configure logging with more detail
logging.basicConfig(filename="predictions.log", level=logging.DEBUG, format="%(asctime)s - %(levelname)s - %(message)s")

app = FastAPI(title="Improved AI Text Detector")

# Enable GPU if available, else use CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch.manual_seed(42)

# Load classifier models
english_detectors = [
    pipeline("text-classification", model="Hello-SimpleAI/chatgpt-detector-roberta", device=device if device.type == "cuda" else -1, truncation=True, max_length=512),
    pipeline("text-classification", model="openai-community/roberta-large-openai-detector", device=device if device.type == "cuda" else -1, truncation=True, max_length=512)
]
arabic_detector = pipeline("text-classification", model="sabaridsnfuji/arabic-ai-text-detector", device=device if device.type == "cuda" else -1, truncation=True, max_length=512)

# Load perplexity models
ppl_english = {
    "tokenizer": AutoTokenizer.from_pretrained("gpt2"),
    "model": AutoModelForCausalLM.from_pretrained("gpt2").to(device)
}
ppl_arabic = {
    "tokenizer": AutoTokenizer.from_pretrained("aubmindlab/aragpt2-base"),
    "model": AutoModelForCausalLM.from_pretrained("aubmindlab/aragpt2-base").to(device)
}

def detect_language(text: str) -> str:
    """Detect if text is Arabic or English based on Unicode character ranges."""
    arabic_chars = len(re.findall(r'[\u0600-\u06FF]', text))
    latin_chars = len(re.findall(r'[A-Za-z]', text))
    total_chars = arabic_chars + latin_chars
    if total_chars == 0:
        return 'en'
    arabic_ratio = arabic_chars / total_chars
    return 'ar' if arabic_ratio > 0.5 else 'en'

def calculate_burstiness(text: str) -> float:
    """Calculate burstiness (std/mean of sentence lengths) to bias toward human text."""
    sentences = re.split(r'[.!?]', text)
    lengths = [len(s.split()) for s in sentences if s]
    return np.std(lengths) / (np.mean(lengths) + 1e-6) if lengths else 0

def calculate_ttr(text: str) -> float:
    """Calculate type-token ratio (lexical diversity) to bias toward human text."""
    words = text.split()
    if not words:
        return 0
    unique_words = len(set(words))
    total_words = len(words)
    return unique_words / total_words

def clean_text(text: str, language: str) -> str:
    """Clean text by removing special characters and normalizing spaces. Skip lowercase for Arabic."""
    text = re.sub(r'\s+', ' ', text)
    text = re.sub(r'[^\w\s.,!?]', '', text)
    text = text.strip()
    if language == 'en':
        text = text.lower()
    return text

def get_classifier_score(text: str, detector) -> float:
    """Get classifier probability for AI label."""
    result = detector(text, truncation=True, max_length=512)[0]
    score = result['score']
    return score if result['label'] in ['AI', 'Fake'] else 1 - score

def get_perplexity(text: str, tokenizer, model) -> float:
    """Calculate perplexity using a language model."""
    encodings = tokenizer(text, return_tensors="pt", truncation=True, max_length=512).to(device)
    max_length = model.config.n_positions
    stride = 512
    seq_len = encodings.input_ids.size(1)

    nlls = []
    prev_end_loc = 0
    for begin_loc in range(0, seq_len, stride):
        end_loc = min(begin_loc + stride, seq_len)
        trg_len = end_loc - prev_end_loc
        input_ids = encodings.input_ids[:, begin_loc:end_loc].to(device)
        target_ids = input_ids.clone()
        target_ids[:, :-trg_len] = -100

        with torch.no_grad():
            outputs = model(input_ids, labels=target_ids)
            neg_log_likelihood = outputs.loss * trg_len

        nlls.append(neg_log_likelihood)
        prev_end_loc = end_loc

        if end_loc == seq_len:
            break

    ppl = torch.exp(torch.stack(nlls).sum() / end_loc if nlls else torch.tensor(0)).item()
    return ppl

def calculate_weighted_score(clf_score: float, ppl: float, burstiness: float, ttr: float, detected_lang: str) -> float:
    """Calculate a weighted score combining classifier and features."""
    ppl_norm = min(ppl / 200, 1.0)  # Normalize perplexity (cap at 200)
    burstiness_norm = min(burstiness / (2.0 if detected_lang == 'en' else 1.5), 1.0)  # Normalize burstiness
    ttr_norm = max(0.1 / max(ttr, 0.01), 1.0)  # Normalize TTR (inverse, cap at 0.1)
    feature_score = (ppl_norm + burstiness_norm + ttr_norm) / 3  # Average feature score
    return 0.6 * clf_score + 0.4 * feature_score  # Weight classifier higher

def split_text(text: str, max_chars: int = 5000) -> list:
    """Split text into chunks of max_chars, preserving sentence boundaries."""
    sentences = re.split(r'(?<=[.!?])\s+', text)
    chunks = []
    current_chunk = ""
    for sentence in sentences:
        if len(current_chunk) + len(sentence) <= max_chars:
            current_chunk += sentence + " "
        else:
            if current_chunk:
                chunks.append(current_chunk.strip())
            current_chunk = sentence + " "
    if current_chunk:
        chunks.append(current_chunk.strip())
    return chunks

class TextInput(BaseModel):
    text: str

    @validator("text")
    def validate_text(cls, value):
        """Validate input text for minimum length and content."""
        word_count = len(value.split())
        if word_count < 50:
            raise ValueError(f"Text too short ({word_count} words). Minimum 50 words required.")
        if not re.search(r'[\w]', value):
            raise ValueError("Text must contain alphabetic characters.")
        return value

@app.post("/detect")
def detect(input_text: TextInput):
    detected_lang = detect_language(input_text.text)
    note_lang = f"Detected language: {'Arabic' if detected_lang == 'ar' else 'English'}"
    cleaned_text = clean_text(input_text.text, detected_lang)
    burstiness = calculate_burstiness(cleaned_text)
    ttr = calculate_ttr(cleaned_text)
    ppl_model = ppl_english if detected_lang == 'en' else ppl_arabic
    ppl = get_perplexity(cleaned_text, ppl_model["tokenizer"], ppl_model["model"])
    note_features = f"Burstiness: {burstiness:.2f} (high suggests human), TTR: {ttr:.2f} (low suggests human), Perplexity: {ppl:.2f} (high suggests human)"

    # Select appropriate models
    detectors = english_detectors if detected_lang == 'en' else [arabic_detector]
    is_ensemble = detected_lang == 'en'

    # Thresholds for human classification
    ppl_threshold = 150  # Increased from 100
    burstiness_threshold = 1.7 if detected_lang == 'en' else 1.2  # Increased from 1.5/1.0
    ttr_threshold = 0.08  # Decreased from 0.10

    if len(cleaned_text) > 10000:
        chunks = split_text(cleaned_text, max_chars=5000)
        labels = []
        clf_scores = []
        ppls = []

        for chunk_idx, chunk in enumerate(chunks):
            chunk_labels = []
            chunk_clf_scores = []
            for det_idx, detector in enumerate(detectors):
                clf_score = get_classifier_score(chunk, detector)
                label = "AI" if clf_score >= 0.90 else "Human" if clf_score < 0.60 else "Uncertain"  # Adjusted from 0.95
                chunk_labels.append(label)
                chunk_clf_scores.append(clf_score)
                logging.debug(f"Chunk {chunk_idx}, Model {det_idx}: Label={label}, Classifier Score={clf_score:.4f}")
            chunk_ppl = get_perplexity(chunk, ppl_model["tokenizer"], ppl_model["model"])
            chunk_final_label = Counter(chunk_labels).most_common(1)[0][0]
            avg_clf_score = np.mean(chunk_clf_scores)

            # Count human-like features
            human_features = sum([
                chunk_ppl > ppl_threshold,
                burstiness > burstiness_threshold,
                ttr < ttr_threshold
            ])
            feature_note = f"Human-like features: {human_features}/3 (PPL={chunk_ppl:.2f}, Burstiness={burstiness:.2f}, TTR={ttr:.2f})"

            # Calculate weighted score
            weighted_score = calculate_weighted_score(avg_clf_score, chunk_ppl, burstiness, ttr, detected_lang)
            chunk_final_label = "AI" if weighted_score >= 0.7 else "Human" if weighted_score < 0.4 else "Uncertain"
            # Require all 3 features to override to Human
            if chunk_final_label == "Uncertain" or any(l == "Human" for l in chunk_labels):
                if human_features == 3:
                    chunk_final_label = "Human"
            elif chunk_final_label == "AI" and avg_clf_score < 0.90 and human_features == 3:
                chunk_final_label = "Human"

            labels.append(chunk_final_label)
            clf_scores.append(avg_clf_score)
            ppls.append(chunk_ppl)
            logging.debug(f"Chunk {chunk_idx} Final: Label={chunk_final_label}, Avg Classifier Score={avg_clf_score:.4f}, Weighted Score={weighted_score:.4f}, Perplexity={chunk_ppl:.2f}, {feature_note}")

        label_counts = Counter(labels)
        final_label = label_counts.most_common(1)[0][0]
        avg_weighted_score = sum(calculate_weighted_score(clf_scores[i], ppls[i], burstiness, ttr, detected_lang) for i in range(len(clf_scores))) / len(clf_scores) if clf_scores else 0.0
        final_label = "AI" if avg_weighted_score >= 0.7 else "Human" if avg_weighted_score < 0.4 else "Uncertain"
        if final_label == "Uncertain" or any(l == "Human" for l in labels):
            human_features = sum([
                any(ppl > ppl_threshold for ppl in ppls),
                burstiness > burstiness_threshold,
                ttr < ttr_threshold
            ])
            if human_features == 3:
                final_label = "Human"

        avg_clf_score = sum(clf_scores) / len(clf_scores) if clf_scores else 0.0
        avg_ppl = sum(ppls) / len(ppls) if ppls else 0.0
        logging.info(f"Language: {detected_lang} | Text Length: {len(cleaned_text)} | Chunks: {len(chunks)} | Prediction: {final_label} | Avg Classifier Score: {avg_clf_score:.4f} | Avg Perplexity: {avg_ppl:.2f} | {note_features}")
        return {
            "prediction": final_label,
            "classifier_score": round(avg_clf_score, 4),
            "perplexity": round(avg_ppl, 2),
            "note": f"{note_lang}. Text was split into {len(chunks)} chunks due to length > 10,000 characters. {note_features}. Weighted Score={avg_weighted_score:.4f}.",
            "chunk_results": [
                {"chunk": chunk[:50] + "...", "label": labels[i], "classifier_score": clf_scores[i], "perplexity": ppls[i], "burstiness": burstiness, "ttr": ttr}
                for i, chunk in enumerate(chunks)
            ]
        }
    else:
        if is_ensemble:
            clf_scores = []
            labels = []
            for det_idx, detector in enumerate(detectors):
                clf_score = get_classifier_score(cleaned_text, detector)
                label = "AI" if clf_score >= 0.90 else "Human" if clf_score < 0.60 else "Uncertain"  # Adjusted from 0.95
                labels.append(label)
                clf_scores.append(clf_score)
                logging.debug(f"Model {det_idx}: Label={label}, Classifier Score={clf_score:.4f}")
            label_counts = Counter(labels)
            final_label = label_counts.most_common(1)[0][0]
            avg_clf_score = sum(clf_scores) / len(clf_scores) if clf_scores else 0.0

            # Count human-like features
            human_features = sum([
                ppl > ppl_threshold,
                burstiness > burstiness_threshold,
                ttr < ttr_threshold
            ])
            feature_note = f"Human-like features: {human_features}/3 (PPL={ppl:.2f}, Burstiness={burstiness:.2f}, TTR={ttr:.2f})"

            # Calculate weighted score
            weighted_score = calculate_weighted_score(avg_clf_score, ppl, burstiness, ttr, detected_lang)
            final_label = "AI" if weighted_score >= 0.7 else "Human" if weighted_score < 0.4 else "Uncertain"
            # Require all 3 features to override to Human
            if final_label == "Uncertain" or any(l == "Human" for l in labels):
                if human_features == 3:
                    final_label = "Human"
            elif final_label == "AI" and avg_clf_score < 0.90 and human_features == 3:
                final_label = "Human"

            note = f"{note_lang}. Ensemble used: {len(detectors)} models. {note_features}. {feature_note}. Weighted Score={weighted_score:.4f}."
            if 0.60 <= avg_clf_score < 0.90:
                note += " Warning: Close to threshold, result may be uncertain."
            logging.info(f"Language: {detected_lang} | Text Length: {len(cleaned_text)} | Prediction: {final_label} | Avg Classifier Score: {avg_clf_score:.4f} | Perplexity: {ppl:.2f} | {note_features} | {feature_note}")
        else:
            clf_score = get_classifier_score(cleaned_text, arabic_detector)
            final_label = "AI" if clf_score >= 0.90 else "Human" if clf_score < 0.60 else "Uncertain"  # Adjusted from 0.95
            # Count human-like features
            human_features = sum([
                ppl > ppl_threshold,
                burstiness > burstiness_threshold,
                ttr < ttr_threshold
            ])
            feature_note = f"Human-like features: {human_features}/3 (PPL={ppl:.2f}, Burstiness={burstiness:.2f}, TTR={ttr:.2f})"

            # Calculate weighted score
            weighted_score = calculate_weighted_score(clf_score, ppl, burstiness, ttr, detected_lang)
            final_label = "AI" if weighted_score >= 0.7 else "Human" if weighted_score < 0.4 else "Uncertain"
            # Require all 3 features to override to Human
            if final_label == "Uncertain" or final_label == "Human":
                if human_features == 3:
                    final_label = "Human"
            elif final_label == "AI" and clf_score < 0.90 and human_features == 3:
                final_label = "Human"

            note = f"{note_lang}. {note_features}. {feature_note}. Weighted Score={weighted_score:.4f}."
            if 0.60 <= clf_score < 0.90:
                note += " Warning: Close to threshold, result may be uncertain."
            logging.info(f"Language: {detected_lang} | Text Length: {len(cleaned_text)} | Prediction: {final_label} | Classifier Score: {clf_score:.4f} | Perplexity: {ppl:.2f} | {note_features} | {feature_note}")
        return {
            "prediction": final_label,
            "classifier_score": round(avg_clf_score, 4),
            "perplexity": round(ppl, 2),
            "note": note
        }