Adversarial training, in which a network is trained on adversarial examples, is one of the few defenses against adversarial attacks that withstands strong attacks. Unfortunately, the high cost of generating strong adversarial examples makes standard adversarial training impractical on large-scale problems like ImageNet. We present an algorithm that eliminates the overhead cost of generating adversarial examples by recycling the gradient information computed when updating model parameters.Our "free" adversarial training algorithm achieves comparable robustness to PGD adversarial training on the CIFAR-10 and CIFAR-100 datasets at negligible additional cost compared to natural training, and can be 7 to 30 times faster than other strong adversarial training methods. Using a single workstation with 4 P100 GPUs and 2 days of runtime, we can train a robust model for the large-scale ImageNet classification task that maintains 40% accuracy against PGD attacks. The code is available at https://github.com/ashafahi/free_adv_train.

SUMMARY Bowel function was assessed in 51 subjects: 10 women and seven men who habitually consumed an omnivorous, vegetarian, or vegan diet. The subjects on these diets had a mean intake of fibre of 23 g, 37 g, and 47 g respectively. Mean transit times were variable and not significantly different between the groups. Vegans, however, had a greater frequency of defecation and passed softer stools. All measurements of bowel function were significantly correlated with total dietary fibre. As dietary fibre increased mean transit time decreased, stool frequency increased and the stools became softer. Men produced a greater quantity of softer, less formed faeces than women. During the luteal phase of the menstrual cycle women excreted harder stools and had a significantly longer mean transit time. The finding that mean transit time was more highly correlated with faecal form than any of the other bowel function measurements could be of practical importance. Critical examination of the methodologies used in these various studies, however, limits interpretation of the results in relation to the 'dietary fibre hypothesis'. For example, duration of diet was not always stated5 and dietary fibre intake was only measured in one of the studies.3 In two of the studies the health of subjects was uncertain.1 4 The sex distribution was either not stated,6 or uneven3 4 5 and in one of the studies all the subjects were men.2 The exact age distribution was not certain in some of the studies1 3 and the age ranges differed markedly in Holm and Hansen's2 subjects. The period of time for which stools were collected in relation to specific measurements of bowel function was probably satisfactory in only three of the studies. 1-3 The method used for the measurement of transit time in two of the studies has been superseded1 2 and of possible significance is the fact that none of the studies took account of stool consistency Address for correspondence: Professor J W T Dickerson, Department of Biochemistry, University of Surrey, Guildford, Surrey GU2 5XH. Received for publication 3 June 1985 or faecal form. Only two of the studies were undertaken in the United Kingdom,1 3 so that geographical factors may limit interpretation.An attempt has been made in the present study to overcome the major uncertainties outlined above. Methods SUBJECTSThe subjects-were all free from overt symptoms of disease, none had undergone major surgery during the preceding six months and none were pregnant or lactating. All subjects were accustomed to their diets. Omnivores had been so for life and the vegetarians and vegans had followed their respective diets for 31% (range 3-35%) and 15% (range 1-41%) of their lives respectively. Most of the vegans had been vegetarians previously. There were 10 women and seven men in each dietary group. The mean ages of the omnivores, vegetarians, and vegans were 35 years (range 11-56 years), 34 years (range 10-71 years) and 31 years (range 5-63 years) respectively. Most of the subjects were in social class 2.7 Menstruating wome...

The fatty acid composition of erythrocytes, of plasma choline phosphoglycerides, and of adipose tissue, serum cholesterol, triglyceride and vitamin B12 concentrations, weights, heights and skinfold thickness were determined on 22 vegans and 22 age and sex matched omnivore controls. The fatty acid composition of breast milk from four vegan and four omnivore control mothers, and of erythrocytes from three infants breast fed by vegan mothers and six infants breast fed by omnivore control mothers was determined. The proportions of linoleic acid and its long-chain derivatives were higher, the proportion of the long-chain derivatives of alpha-linolenic acid was lower, and the ratio of 22:5omega3/22:6omega3 was greater in the tissues of the vegans and infants breast-fed by vegans than in controls; the most marked differences were in the proportions of linoleic (18:2omega6) and docosahexenoic (22:6omega3) acids. Weights, skinfold thickness, serum vitamin B12, cholesterol and triglyceride concentrations were less in vegans than in controls. The difference in serum cholesterol concentration was most marked. It is concluded that a vegan-type diet may be the one of choice in the treatment of ischemic heart disease, angina pectoris, and certain hyperlipidemias.

No abstract